Model rankings

Every model, ranked on quality, speed and cost.

Three things decide a model: how well it thinks, how fast it answers, and what one finished task costs. Every model we serve, plus what's coming, on all three.

Start here

What are you building?

The best model for one job is rarely the best for the next. Find the row that sounds like your week.

Agentic coding

Many steps, many files, and the model has to keep its own plan straight for an hour. A one-shot index only half predicts that, so run your own repo through the top two.

Pick

GPT 5.6 Sol orClaude Opus 5

On a budget

DeepSeek V4 Pro $0.27/task at top effort
The genuinely hard problems

Research, analysis and anything where being wrong is expensive.

Pick

Claude Opus 5 orClaude Fable 5

On a budget

Kimi K3 $0.84/task at top effort
Anything a person waits on

Chat, voice, autocomplete. The first token has to land before attention does.

Pick

Gemini 3.7 Flash soonorGPT 5.6 Luna
Documents, images, audio, video

One request carrying a whole corpus of mixed media, reasoned over as a unit.

Pick

Gemini 3.7 Flash soonorMuse Spark 1.2 soonorGemini 3.6 Flash
High volume, tight budget

Classification, extraction, tagging — millions of calls where the bill is the constraint.

Pick

GPT 5.6 Luna orDeepSeek V4 Flash soon
Weights you can run yourself

Data that cannot leave your environment, or volume that beats renting.

Pick

Kimi K3 orDeepSeek V4 Pro

On a budget

DeepSeek V4 Flash soon $0.11/task at top effort

Starting cheap? mindshub_air is our low-cost tier — an alias pointed at the best-scoring model at the cheap end of the rate card (today GPT 5.6 Luna), with 5M tokens a month included. No web search, no reasoning dial, not a frontier model — the name is the tier. It has no row here because it is an alias, not a model.

Leaderboard

Every model, every axis.

Sort by what you care about, filter to what you can use, and expand a row for the parts that don't fit in a cell.

Large language models ranked by quality index, with coding index, output speed, latency, measured cost per task at top reasoning effort, price per million tokens, context window and weight availability.

Anthropic `opus` Tops the intelligence index; built for long agentic runs	63 Frontier	72.4	55 tok/sec	3.2s first token	$2.34	$5 / $25 our rate	1M	Closed
Pick it when The task is hard, runs for many steps, or the output carries risk that someone has to sign off on. Skip it when You're paying by the token for routine work — it keeps chewing long after a cheaper model would have stopped. Inputs TextImages API alias `opus` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `lowmediumhighxhighmax` We send high unless you ask for another. Quality and Cost/task above were measured at max, this model's top setting, so this row's board position is a ceiling rather than its out-of-the-box behaviour: at our default you pay less, and on hard work you may score lower. Worth knowing The top of its dial is a plateau, not a peak. This row quotes max; one notch down at xhigh Artificial Analysis measures the same 63 for $1.80 a task instead of $2.34. Read the score as a ceiling and the cost as a worst case.
Anthropic `fable` Anthropic's top tier for the hardest problems	62 Frontier	75	70 tok/sec	3.6s first token	$3.14	$10 / $50 our rate	1M	Closed
Pick it when Nothing else has cleared the bar and the answer is worth a premium rate. Skip it when Opus 5 sits one index point away at half the price — start there and only move up if it falls short. Inputs TextImages API alias `fable` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `lowmediumhighxhighmax` We send high unless you ask for another. Quality and Cost/task above were measured at max, this model's top setting, so this row's board position is a ceiling rather than its out-of-the-box behaviour: at our default you pay less, and on hard work you may score lower.
OpenAI `gpt` Flagship all-rounder, and first on the coding index	61 Frontier	75.6	73 tok/sec	2.0s first token	$1.01	$5 / $30 our rate	1M	Closed
Pick it when You want one provider that is competent at everything, with the deepest tooling ecosystem around it. Skip it when Price is the binding constraint — GPT 5.6 Terra scores four points lower for about half the cost per finished task. Inputs TextImages API alias `gpt` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `nonelowmediumhighxhighmax` We send medium unless you ask for another. Quality and Cost/task above were measured at max, this model's top setting, so this row's board position is a ceiling rather than its out-of-the-box behaviour: at our default you pay less, and on hard work you may score lower.
Zhipu coming soon Frontier-class coding at a low API price	60 Near-frontier	65	84 tok/sec	1.6s first token	$0.68	$1.40 / $4.40 list price	1M	API only
Pick it when You want strong coding and security review at a fraction of frontier prices, and renting is fine. Skip it when You need multimodal input, or weights you can host today. Inputs Text Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Worth knowing Zhipu has not released the weights, so this is an API you rent rather than a model you can host. Price shown Zhipu's own published list price, not ours — we don't serve this model yet.
Moonshot AI `kimi` Long autonomous coding runs you can self-host	60 Near-frontier	60.2	39 tok/sec	3.4s first token	$0.84	$3 / $15 our rate	1M	Open Moonshot licence
Pick it when You want the strongest model you can genuinely download and run yourself. Skip it when The API bill is the point — GLM is cheaper for similar work. Inputs TextImages API alias `kimi` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort Single setting — no dial. Nothing to turn down, so the columns above are what you get. That also makes it predictable to budget. Licence Moonshot licence Served by Fireworks AI Moonshot AI makes it; we buy the inference from Fireworks AI, which is the name the rate card shows. Worth knowing The licence is Moonshot's own, not MIT or Apache. Self-hosting and fine-tuning are fine; reselling access at scale needs a read.
SpaceXAI `grok` Reasoning with access to the live web	60 Near-frontier	60	61 tok/sec	1.2s first token	$1.23	$2 / $6 our rate	500K	Closed
Pick it when You need information from this morning in the same call as the reasoning. Skip it when Your agents live in a terminal, or your prompts are enormous — the window is half what most of this board offers. Inputs TextImages API alias `grok` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `lowmediumhighxhigh` We send high unless you ask for another. Quality and Cost/task above were measured at xhigh, this model's top setting, so this row's board position is a ceiling rather than its out-of-the-box behaviour: at our default you pay less, and on hard work you may score lower. Worth knowing Its dial stops paying off at the top. This row quotes xhigh, the highest setting Artificial Analysis measures, and one notch down at high the same model scores a point higher for a quarter less per task. What the top setting buys is four tokens a second.
Alibaba coming soon The open checkpoint that matches the Qwen flagship	58 Near-frontier	56	24 tok/sec	4.1s first token	$0.81	$2 / $6 list price	984K	Open Qwen licence (commercial use restricted)
Pick it when You need the weights, and text-only input plus a restrictive licence are acceptable. Skip it when You are reselling access, or your inputs are not text. Inputs Text Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Licence Qwen licence (commercial use restricted) Worth knowing Artificial Analysis costs it at Alibaba's own $2 / $6 — the same rate card as the proprietary Max — but the weights are downloadable, so your actual bill is whatever your host or your hardware charges. Price shown Alibaba's own published list price, not ours — we don't serve this model yet.
Alibaba coming soon Multimodal and multilingual work on a budget	58 Near-frontier	57	21 tok/sec	3.9s first token	$0.91	$2 / $6 list price	1M	Closed
Pick it when You want a capable multimodal API with a million-token window and strong non-English coverage. Skip it when Speed or cost per task is the priority — it is slow and verbose, and keeps almost none of its token discount. Inputs TextImagesVideo Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Worth knowing The Price column puts it at a fifth of GPT 5.6 Sol per output token, and it still comes out only 1.1x cheaper per finished task — it is slow and it thinks at length, and both show up on the invoice. Price shown Alibaba's own published list price, not ours — we don't serve this model yet.
Zhipu coming soon A high score per dollar, on MIT-licensed weights	57 Strong	62	50 tok/sec	1.5s first token	$0.09	$0.15 / $0.50 list price	1M	Open MIT
Pick it when You want a high score per dollar and MIT weights, and you can absorb a slow first token. Skip it when You are waiting on a person — it is one of the slower models here. Inputs TextImages Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Licence MIT Worth knowing The score comes at a price in patience: Artificial Analysis measures it as slow and very verbose, so a finished answer takes longer than its cost per task suggests. First natively multimodal model in the GLM-5 line — 320B total parameters, 18B active. Price shown Zhipu's own published list price, not ours — we don't serve this model yet.
Meta coming soon Cheap closed model that powers the free Meta AI	57 Strong	54	158 tok/sec	0.7s first token	$0.40	$1.25 / $4.25 list price	1M	Closed
Pick it when You want a fast, capable, every-input model and Meta is an acceptable processor of your data. Skip it when You need weights — the Muse family is closed — or your data cannot go to Meta. Inputs TextImagesAudioVideo Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Worth knowing Meta publishes a "contributor" tier at roughly a fifteenth of the standard rate in exchange for training on your prompts and replies. Fine for hobby projects, an easy no for anything confidential. Price shown Meta's own published list price, not ours — we don't serve this model yet.
OpenAI `gpt-terra` Balanced mid-tier for everyday work	57 Strong	66	109 tok/sec	1.3s first token	$0.53	$2 / $12 our rate	1M	Closed
Pick it when You want the sensible default: strong enough for most work, and it keeps almost all of its discount on the invoice. Skip it when The last few index points decide the outcome. Inputs TextImages API alias `gpt-terra` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `nonelowmediumhighxhighmax` We send medium unless you ask for another. Quality and Cost/task above were measured at max, this model's top setting, so this row's board position is a ceiling rather than its out-of-the-box behaviour: at our default you pay less, and on hard work you may score lower.
Google coming soon Fast multimodal work at a promotional price	56 Strong	72	362 tok/sec	0.5s first token	$0.40	$0.75 / $3.75 list price	1M	Closed
Pick it when Inputs are large or mixed-media — PDFs, images, audio, video — and you want them read quickly. Skip it when You need frontier-level reasoning, or weights you can host yourself. Inputs TextImagesAudioVideo Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Price shown Google's own published list price, not ours — we don't serve this model yet. Promotional through December 31, 2026.
SpaceXAI `grok-4-5` The previous Grok, and now the better-value one	56 Strong	56	56 tok/sec	1.4s first token	$0.43	$2 / $6 our rate	500K	Closed
Pick it when You want most of 4.6 for much less per task, or you have production traffic validated against this version. Skip it when You need the four index points 4.6 adds. Inputs TextImages API alias `grok-4-5` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `lowmediumhigh` We send high unless you ask for another, and that is this model's top setting — the one Quality and Cost/task above were measured at. Turning it down is the saving on offer here, not a surprise on the invoice.
Anthropic `sonnet` Balanced everyday workhorse	55 Strong	70	92 tok/sec	2.4s first token	$1.72	$2 / $10 our rate	1M	Closed
Pick it when You want one model that drafts, summarizes, analyzes and codes without hitting a wall. Skip it when Volume is the point. It costs more per finished task than its rate card suggests. Inputs TextImages API alias `sonnet` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `lowmediumhighmax` We send high unless you ask for another. Quality and Cost/task above were measured at max, this model's top setting, so this row's board position is a ceiling rather than its out-of-the-box behaviour: at our default you pay less, and on hard work you may score lower. Worth knowing Three times cheaper than GPT 5.6 Sol per output token and about 70% more per finished task — the plainest case on this board that a rate card does not predict a bill. Its streaming rate is the one on this board that moves most with effort: 92 tok/s at the top setting these figures come from, against 68 at the high setting we send by default.
Meta `muse-spark` Every input type, at an open-model rate	54 Solid	52	150 tok/sec	0.8s first token	$0.44	$1.25 / $4.25 our rate	1M	Closed
Pick it when You want every input type at a low rate and a fast first token. Skip it when 1.2 is three index points better on the same rate card. Inputs TextImagesAudioVideo API alias `muse-spark` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort Single setting — no dial. Nothing to turn down, so the columns above are what you get. That also makes it predictable to budget.
OpenAI `gpt-codex` Coding specialist for repo-scale changes	54 Solid	68	136 tok/sec	1.7s first token	—	$1.75 / $14 our rate	400K	Closed
Pick it when The job is software, across many files, and you want a model tuned for that rather than for general reasoning. Skip it when The work is not code. Its general index score sits below its coding score for a reason. Inputs TextImages API alias `gpt-codex` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `lowmediumhighxhigh` We send medium unless you ask for another. Quality and Cost/task above were measured at xhigh, this model's top setting, so this row's board position is a ceiling rather than its out-of-the-box behaviour: at our default you pay less, and on hard work you may score lower. Worth knowing Artificial Analysis scores this version only provisionally — on an older component set that omits all three of the benchmarks they run on every fully scored row (GDPval-AA v2, Terminal-Bench v2.1 and t3-Banking) — and publishes no cost per task for it, as they do for none of their provisional scores. Their figure is not on the same scale as the rest of this column, so both published columns here are ours.
DeepSeek `deepseek` Open-weight reasoning and coding at a budget rate	53 Solid	58.6	68 tok/sec	2.6s first token	$0.27	$1.74 / $3.48 our rate	1M	Open MIT
Pick it when Volume is high, the budget is tight, and you want permissive weights as an exit route. Skip it when You need multimodal input or the last few points of quality. Inputs Text API alias `deepseek` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `nonelowmediumhighxhighmax` We send high unless you ask for another. Quality and Cost/task above were measured at max, this model's top setting, so this row's board position is a ceiling rather than its out-of-the-box behaviour: at our default you pay less, and on hard work you may score lower. Licence MIT Served by Fireworks AI DeepSeek makes it; we buy the inference from Fireworks AI, which is the name the rate card shows.
Zhipu `glm` Open weights with the GLM coding profile	53 Solid	62	69 tok/sec	1.8s first token	$0.44	$1.40 / $4.40 our rate	1M	Open
Pick it when You want the GLM coding profile and weights you can host, and can spare the points 5.3 adds. Skip it when 5.3 is seven index points better on the same rate card, if renting is fine. Inputs Text API alias `glm` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `highmax` We send max unless you ask for another, and that is this model's top setting — the one Quality and Cost/task above were measured at. Turning it down is the saving on offer here, not a surprise on the invoice. Served by Fireworks AI Zhipu makes it; we buy the inference from Fireworks AI, which is the name the rate card shows. Worth knowing Artificial Analysis now lists these weights as open, while 5.3 is still API-only — so the older release is the one you can host. The licence name is not recorded here; read Zhipu's terms before you build on it.
OpenAI `gpt-luna` Fast and cheap for high-volume work	52 Solid	48	131 tok/sec	0.5s first token	$0.05	$0.20 / $1.20 our rate	1M	Closed
Pick it when Volume is high, tasks are routine, and the bill matters more than the last few points of quality. Skip it when The work is genuinely hard. It scores alongside good open models, not alongside the frontier. Inputs TextImages API alias `gpt-luna` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `nonelowmediumhighxhighmax` We send medium unless you ask for another. Quality and Cost/task above were measured at max, this model's top setting, so this row's board position is a ceiling rather than its out-of-the-box behaviour: at our default you pay less, and on hard work you may score lower.
DeepSeek coming soon Low-cost, MIT-licensed, and you can host it yourself	52 Solid	52	122 tok/sec	0.9s first token	$0.11	$0.44 / $1.32 list price	1M	Open MIT
Pick it when High-volume routine work where you want the option to stop renting and self-host. Skip it when Quality is the constraint rather than cost. Inputs Text Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Licence MIT Price shown DeepSeek's own published list price, not ours — we don't serve this model yet.
Google `gemini-flash` Fast, high-volume everyday tasks	52 Solid	68.8	203 tok/sec	0.6s first token	$0.34	$0.75 / $3.75 our rate	1M	Closed
Pick it when You want multimodal input and a million-token window at a throughput price. Skip it when 3.7 Flash scores four index points better and is stronger at code. Inputs TextImagesAudioVideo API alias `gemini-flash` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `mediumhigh` We send medium unless you ask for another — the lowest setting it has. Quality and Cost/task above were measured at high, this model's top setting, so this row's board position is a ceiling rather than its out-of-the-box behaviour: at our default you pay less, and on hard work you may score lower.
OpenAI `gpt-mini` Small and cheap for routine automation	48 Light	50	150 tok/sec	0.7s first token	$0.19	$0.75 / $4.50 our rate	400K	Closed
Pick it when Classification, extraction, tagging and other high-frequency jobs with a narrow definition of correct. Skip it when Anything open-ended. Luna is both cheaper and better at this point in the line-up. Inputs TextImages API alias `gpt-mini` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `nonelowmediumhighxhigh` We send none unless you ask for another — the lowest setting it has, and no reasoning tokens at all. Quality and Cost/task above were measured at xhigh, this model's top setting, so this row's board position is a ceiling rather than its out-of-the-box behaviour: at our default you pay less, and on hard work you may score lower.
Google `gemini` Huge documents, images, audio and video	48 Light	68.8	117 tok/sec	1.6s first token	$0.33	$2 / $12 our rate	1M	Closed
Pick it when A single request has to carry a whole corpus of mixed media, and it has to be reasoned over rather than skimmed. Skip it when You are paying by the token — the Flash line reads the same inputs at a third of the rate card, and 3.7 Flash outscores this by eight points. Inputs TextImagesAudioVideo API alias `gemini` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `minimallowmediumhigh` We send high unless you ask for another, and that is this model's top setting — the one Quality and Cost/task above were measured at. Turning it down is the saving on offer here, not a surprise on the invoice.
Moreh coming soon Korea's sovereign open model, free to download	47 Light	—	—	—	—	— host-dependent	262K	Open Motif research licence (non-commercial)
Pick it when You want a strong open checkpoint for research or internal evaluation and text-only input is fine. Skip it when It is going anywhere near a product — the licence rules out commercial use. Inputs Text Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Licence Motif research licence (non-commercial) Worth knowing Four columns are blank because nobody has published them. The weights are free and there is no priced first-party endpoint, so Artificial Analysis reports its cost per task as unknown — the bill is whatever your own hardware or host costs. It is also very verbose: 260M output tokens across the Intelligence Index against a 100M median, which is the shape of a model that thinks in long answers. Price shown Moreh's own published list price, not ours — we don't serve this model yet.
MiniMax coming soon A quietly strong multimodal all-rounder, very cheap	45 Light	44	114 tok/sec	0.9s first token	$0.14	$0.30 / $1.20 list price	1M	Open MiniMax Community (commercial use restricted)
Pick it when You want a million-token window and image input at the bottom of the price range. Skip it when You are building a commercial product on the weights — read the licence first. Inputs TextImages Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Licence MiniMax Community (commercial use restricted) Price shown MiniMax's own published list price, not ours — we don't serve this model yet.
Thinking Machines Lab coming soon Apache 2.0 weights that take audio in	42 Light	—	50 tok/sec	2.9s first token	$0.34	$1 / $4.05 list price	1M	Open Apache 2.0
Pick it when You need a permissive licence with no commercial catch, and audio input in the same model. Skip it when The score is what you are buying — several open models here rank higher for less. Inputs TextImagesAudio Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Licence Apache 2.0 Worth knowing A genuinely unrestricted licence, at a price: several times the cost per task of its own Small sibling for one index point more, and slower than most of this board. 975B total parameters, 41B active. Price shown Thinking Machines Lab's own published list price, not ours — we don't serve this model yet.
Thinking Machines Lab coming soon The same licence and inputs, two thirds smaller	41 Light	—	109 tok/sec	1.7s first token	$0.07	$0.30 / $1.20 list price	1M	Open Apache 2.0
Pick it when You want the Apache 2.0 terms and the multimodal inputs on hardware you can actually afford. Skip it when You need the one index point Inkling adds. Inputs TextImagesAudio Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Licence Apache 2.0 Worth knowing One index point behind Inkling on well under a third of the parameters, twice as fast and a fifth of the cost per task. On this board it is the better of the pair on every axis except the score. Price shown Thinking Machines Lab's own published list price, not ours — we don't serve this model yet.
Alibaba `qwen` Multilingual coverage at a fraction of the flagship rate	39 Basic	55	56 tok/sec	1.5s first token	$0.17	$0.40 / $1.60 our rate	1M	Closed
Pick it when You want Qwen-family behaviour and broad multilingual coverage without paying a flagship rate. Skip it when You need the current flagship — 3.8-Max is nineteen index points ahead. Inputs TextImages API alias `qwen` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort Single setting — no dial. Nothing to turn down, so the columns above are what you get. That also makes it predictable to budget. Served by Fireworks AI Alibaba makes it; we buy the inference from Fireworks AI, which is the name the rate card shows.
OpenAI `gpt-nano` Built for throughput on bulk, simple work	38 Basic	30	220 tok/sec	0.4s first token	$0.05	$0.20 / $1.25 our rate	400K	Closed
Pick it when Millions of tiny calls where latency is the product and the task fits in one sentence. Skip it when Any judgement is required. Inputs Text API alias `gpt-nano` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort `nonelowmediumhighxhigh` We send none unless you ask for another — the lowest setting it has, and no reasoning tokens at all. Quality and Cost/task above were measured at xhigh, this model's top setting, so this row's board position is a ceiling rather than its out-of-the-box behaviour: at our default you pay less, and on hard work you may score lower.
Meta coming soon A small open agent model that runs on one consumer GPU	35 Basic	—	107 tok/sec	0.8s first token	$0.06	$0.32 / $1.35 list price	131K	Open Apache 2.0
Pick it when You want something genuinely local: Apache 2.0 weights, 30B parameters, and a score that still holds up on well-scoped agent steps. Skip it when The work is open-ended. This is the smallest capable thing here, not a substitute for the rows above it. Inputs TextImages Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Licence Apache 2.0 Worth knowing Meta going closed with Muse Spark did not stop it shipping small open models. The 131K window is the shortest on this board, and inputs beyond text and images are not recorded here. Price shown Meta's own published list price, not ours — we don't serve this model yet.
Anthropic `haiku` Fastest Claude, for high-volume simple work	30 Basic	43.4	119 tok/sec	1.1s first token	$0.22	$1 / $5 our rate	200K	Closed
Pick it when You want Claude behaviour and Claude tooling on work that does not need Claude-sized thinking. Skip it when The job needs reasoning headroom — this is the previous generation, and it shows on hard tasks. Inputs TextImages API alias `haiku` The whole integration is this string. It follows model upgrades, so pinning to it survives a version bump. Reasoning effort Single setting — no dial. Nothing to turn down, so the columns above are what you get. That also makes it predictable to budget.
NVIDIA coming soon Built for throughput on the routine steps of an agent run	24 Basic	—	300 tok/sec	1.3s first token	$0.08	$0.07 / $0.22 list price	1M	Open OpenMDW-1.1
Pick it when An agent is grinding through thousands of small, well-defined steps and the wall clock is the constraint. Skip it when Any step needs judgement — this is the lowest score on the board. Inputs Text Reasoning effort Not known yet. Which levels a model accepts comes from our catalog, and this one isn't in it yet — so we can't tell you, rather than there being nothing to tell. Licence OpenMDW-1.1 Worth knowing It wins the accuracy-versus-speed frontier in its size class rather than the accuracy one: 31.6B total parameters, 3.6B active, and among the fastest output on this board. Weights, and a permissive licence, but a score that only suits work you have already decomposed. Price shown NVIDIA's own published list price, not ours — we don't serve this model yet.

Quality & Cost/task: Artificial Analysis's published figures, at each model's top reasoning effort — so both are a ceiling. Expand a row for the level we send.
Coding, Speed & Latency: Speed is Artificial Analysis's published median where they publish one; coding and latency are our own read. Close values are ties, not differences.
Price: Per 1M tokens, in / out. Pure token prices — the 5% fee is a separate line on the bill. Full rate card.
Missing figures: ≈ is our estimate. — means nobody has published one.
Checked: August 27, 2026. Our rates are live from the catalog.

01 · Quality

The top of the board is flat.

The top 6 scorers sit within 3 points of each other, which is close enough to be noise. When quality is that level, speed and cost decide.

A benchmark is a generic picture, not a verdict on your work. It averages somebody else's tasks, and the average is exactly where a model's weak spot hides — one that aces graduate physics can still write clunky support replies. Use these scores to draw a shortlist. Then run your own tasks through the top two or three and keep the one whose answers you would actually send; there's a way to do that in an afternoon below.

One thing the number hides: every entry is a model at a reasoning-effort setting, and labs submit at maximum. So these are ceilings. Turning the dial down moves cost further than switching models does — across Claude Opus 5's five settings, token use on the same suite spans eight-fold. A frontier model turned down usually beats a mid-tier model turned up, on both.

The model-by-model rundown What a benchmark hides Does a new release change your pick? Point a coding agent at any of them

Quality index, whole board. The flat stretch at the top is the point; the drop-off below it is where the index starts deciding things.

Claude Opus 5 63
Claude Fable 5 62
GPT 5.6 Sol 61
GLM-5.3 soon 60
Kimi K3 60
Grok 4.6 60
Qwen3.8 2.4T A95B soon 58
Qwen3.8-Max soon 58
GLM-5.3-Flash soon 57
Muse Spark 1.2 soon 57
GPT 5.6 Terra 57
Gemini 3.7 Flash soon 56
Grok 4.5 56
Claude Sonnet 5 55
Muse Spark 1.1 ≈54
GPT 5.3 Codex ≈54
DeepSeek V4 Pro 53
GLM 5.2 53
GPT 5.6 Luna 52
DeepSeek V4 Flash soon 52
Gemini 3.6 Flash 52
GPT 5.4 Mini ≈48
Gemini 3.1 Pro Preview 48
Motif 3 soon 47
MiniMax-M3 soon 45
Inkling soon 42
Inkling Small soon 41
Qwen3.7 Plus 39
GPT 5.4 Nano ≈38
Muse Glimmer soon 35
Claude Haiku 4.5 30
Nemotron 3.5 Lightning soon 24

Bars run from zero and are relative to the leader — a truncated axis would manufacture a gap that isn't there. Scores within 3 points of each other are a tie.

02 · Performance

Speed is two numbers, not one.

Latency is the wait before an answer starts. Throughput is how fast the rest arrives. Plenty of models are good at one and poor at the other, which is why the chart below shows both.

From the request to a finished answer. Pale is the wait before the first token, solid is the time to stream the rest. A ranking, not a stopwatch: the streaming rate is Artificial Analysis's published median wherever they publish one, the first-token wait is our own read on every row. Answer length changes both the stakes and the order. A 200-token reply lands between 1.1s and 13.4s; at 4,000 tokens the same board runs 12s to 194s, with 24 of 31 rows changing places.

Gemini 3.7 Flash soon 3.3s
Nemotron 3.5 Lightning soon 4.6s
GPT 5.4 Nano 4.9s
Gemini 3.6 Flash 5.5s
Muse Spark 1.2 soon 7.0s
GPT 5.4 Mini 7.4s
Muse Spark 1.1 7.5s
GPT 5.6 Luna 8.1s
GPT 5.3 Codex 9.1s
DeepSeek V4 Flash soon 9.1s
Claude Haiku 4.5 9.5s
MiniMax-M3 soon 9.7s
Muse Glimmer soon 10.1s
Gemini 3.1 Pro Preview 10.1s
GPT 5.6 Terra 10.5s
Inkling Small soon 10.9s
Claude Sonnet 5 13.3s
GLM-5.3 soon 13.5s
GPT 5.6 Sol 15.7s
GLM 5.2 16.3s
DeepSeek V4 Pro 17.3s
Grok 4.6 17.6s
Claude Fable 5 17.9s
Grok 4.5 19.3s
Qwen3.7 Plus 19.4s
Claude Opus 5 21.4s
GLM-5.3-Flash soon 21.5s
Inkling soon 22.9s
Kimi K3 29.0s
Qwen3.8 2.4T A95B soon 45.8s
Qwen3.8-Max soon 51.5s

Wait for the first token Streaming the answer

The spread is 6-fold

GPT 5.4 Nano finishes that answer in about 5 seconds and Kimi K3 takes about 29. Same answer length, same request. Nobody waits 29 seconds for a chat reply, and both of those models are on this board.

Short answers change what to rank on

On a 200-token reply the wait is 25% to 66% of the wall clock, so latency is worth weighting. By a few thousand tokens it is a rounding error and throughput decides alone. The control above the chart switches between the two views.

Reasoning is charged in seconds too

The models that think longest before answering are the ones with the worst latency, because the thinking happens before the first token appears. Turning reasoning effort down shortens the wait and the bill at the same time.

Throughput is a batch-job number

Nobody watches a nightly summarisation run. If no human is waiting, optimise tokens per second and ignore latency entirely — and check whether the provider has a batch endpoint, which often halves the bill for work that is not time-sensitive.

Your numbers will differ

Region, prompt length, cache hit rate and reasoning effort all move these figures. Use them to rank candidates, then measure the two you shortlist against your own traffic.

03 · Cost

The price page stopped predicting the bill.

A price per million tokens is a unit price. It says what one unit costs and nothing about how many units the job will take — and that second number now varies between models far more than the first one does.

Models now decide for themselves how long to think, and you pay for that hidden reasoning exactly as for the visible answer. A model five times cheaper per token can land within pennies of one that isn't. Rank on cost per task instead.

It is somebody else's task, at each model's top effort, priced at the provider's own rate — so treat it as a way to rank candidates, not a forecast of your bill.

Why cheap models often aren't Two levers that cut the bill The full rate card

How much of the discount survives the invoice. Each model against GPT 5.6 Sol: what its published output rate promises, then what its measured cost per task delivers — both at top effort. These 5 keep the least.

Nemotron 3.5 Lightning 90.9× cheaper on paper 12.6× cheaper per task
Grok 4.6 3.3× cheaper on paper 1.2× dearer per task
GLM-5.3-Flash 40.0× cheaper on paper 11.2× cheaper per task
Claude Sonnet 5 2.0× cheaper on paper 1.7× dearer per task
GLM-5.3 4.5× cheaper on paper 1.5× cheaper per task

Both figures come from the same rate card — each provider's own published price, which is what these tasks were costed at, and not always the rate in the Price column above. And the lesson is not that cheap models are a lie: Gemini 3.1 Pro Preview beats its own rate card — 1.7× cheaper on paper, 3.1× cheaper per task. It is that you cannot rank them by reading price pages.

04 · Method

How to actually decide, in an afternoon.

No leaderboard can rank models on your work. This is the shortest thing that can.

Shortlist two or three from the board

Use the task router above, or filter the board to your constraint — open weights, a modality, a latency ceiling. Stop at three.
Collect ten real tasks

With the messy context attached. Not clean samples: the actual ones you ran last week, including the awkward ones.
Run them at the effort you would really use

Not maximum. The reasoning dial moves cost further than switching models does, and every published cost-per-task figure is measured at the top setting.
Read the token counts off the response

Every provider returns a usage object. Multiply input, cached input and output separately — those three rates can differ by an order of magnitude.
Divide by the results you would actually ship

If eight of ten came back usable, your real cost is the total divided by eight. That division is the one no leaderboard can do for you.

One API key reaches every model here, so a bake-off is a string change rather than three integrations. There's a runnable version, an afternoon costing recipe, a guide to running evals, and a free working session if you'd rather do it with an engineer.

FAQ

The questions that decide it.

Which model should I use?

It depends on the job, and the board names a pick for each. Agentic coding: GPT 5.6 Sol or Claude Opus 5 — Sol tops the coding index and Opus tops the intelligence index, and long agentic runs want both. The hardest reasoning: Claude Opus 5, with Kimi K3 as the value option — three index points behind for about a third of the cost per task. GLM-5.3 is cheaper again but is not in the catalog yet. Anything a person waits on: the Gemini Flash line, or GPT 5.6 Luna. Documents, images, audio and video: the Gemini Flash line again. High volume on a budget: GPT 5.6 Luna, at the board's lowest cost per task, $0.05 — matched only by an estimate. Or GLM-5.3-Flash, five index points higher and still under a tenth of a dollar a task, with MIT-licensed weights — though it is not in the MindsHub catalog yet, and not for anything a person is waiting on, because it is one of the slower models on the board. Weights you can host: Kimi K3. Starting cheap rather than choosing on capability: the mindshub_air alias is the low-cost entry point, kept pointed at the best-scoring model at the cheap end of the rate card, with 5M tokens a month at no charge — not a frontier model, and not meant to be.

What does "coming soon" mean on this page?

That the model is benchmarked on the board but is not in the MindsHub catalog yet, so you cannot call it on the API today. It is one small marker on an otherwise identical row, and it clears itself: the board reads the live catalog on every request and binds a row to a model by name, so the moment we start serving one the marker goes, the API alias appears and the price switches to ours — with no change to the page. Until then those rows show the provider's own published list price rather than ours; every model we already serve takes its rates live from our catalog, the same source the pricing page reads.

Why rank on cost per task instead of price per token?

Because a unit price says nothing about how many units the job will take, and that second number now varies between models far more than the first: models decide for themselves how long to think and how many turns to take, and hidden reasoning tokens are billed like visible ones. Grok 4.6 is five times cheaper than GPT 5.6 Sol per output token and costs about 20% more per finished task. The task being measured is one piece of work from Artificial Analysis's suite — a mix of agentic runs, terminal coding, graduate science questions and tool-calling exercises — so treat it as a way to rank candidates rather than as a forecast of your bill. Both price columns on the board are pure token prices with nothing folded in; the platform fee is a separate line on the bill, and the pricing page states it in full.

Where do these benchmark numbers come from?

The page says, column by column, where each figure comes from. The quality index and the cost per task are Artificial Analysis's published measurements, both taken at each model's top reasoning-effort setting, checked on the date shown at the top of the page; a "≈" marks a row they have not covered at that exact model version, where the figure is our own estimate. The coding index and the first-token latency are our own read on every row; the speed column is Artificial Analysis's published median tokens per second wherever they publish one for that exact version, and our read on the four rows they publish no median for. Close values in those three columns are ties rather than measurements. Prices for models we serve are read live from our own catalog.

At what reasoning effort were these numbers measured?

At each model's top setting. Artificial Analysis publishes a separate score and a separate cost for every effort level a model exposes, and labs submit at maximum because that maximises the score — so both published columns on this board are ceilings, and the cost is a worst case. Expand a row to see the levels a model accepts and the one MindsHub sends by default; on most models the default sits below the measured setting, so real bills come in lower. How much lower depends on the model: across Claude Opus 5's five settings, token consumption on the same suite spans roughly eight-fold at one unchanged price per token. We do not publish a per-setting score of our own, because we have not measured one.

Does this page replace the pricing page?

No. The board carries two price dimensions — input and output per million tokens — because that is what a ranking needs, and both are pure token prices with nothing folded into them. Cached input, cache writes, long-prompt surcharge tiers, web search and failover order are all on the pricing page, which is the only place those numbers are authoritative — as is the 5% platform fee, which is a separate line on the bill rather than part of any rate.

Every model, ranked on quality, speed and cost.

What are you building?

Agentic coding

The genuinely hard problems

Anything a person waits on

Documents, images, audio, video

High volume, tight budget

Weights you can run yourself

Every model, every axis.

The top of the board is flat.

Speed is two numbers, not one.

The spread is 6-fold

Short answers change what to rank on

Reasoning is charged in seconds too

Throughput is a batch-job number

Your numbers will differ

The price page stopped predicting the bill.

How to actually decide, in an afternoon.

Shortlist two or three from the board

Collect ten real tasks

Run them at the effort you would really use

Read the token counts off the response

Divide by the results you would actually ship

The questions that decide it.

One key. Every model in the board.