AI Model Observatory · data as of
29 frontier models from 15 labs, compared on intelligence, coding, price, speed, open weights and voice. Every number links to where it came from. Where we could not verify a figure you see a dash — never a guess.
The six verdictsThere is no single winner — there is a winner per job
01Best overall reasoning
Anthropic
Scores 58 at max effort.
Next: Claude Sonnet 5.5 56, Muse Wonder Full 56, Claude Fable 5.1 53.
Caveat At that setting it costs $5.98 per index task and waits 691.7s before the first token.
Highest Artificial Analysis Intelligence Index (independent).
02Best coding
Anthropic
70.6%.
Next: Claude Opus 5.5 66.4%, GPT-6 Astra 57.9%, Claude Fable 5.1 55.8%.
Caveat Every Terminal-Bench figure here is vendor-reported; we could not load an independent coding leaderboard on this date. On a different test, DeepSWE v1.1, Gemini 4 Argon leads at 77.9% (also vendor-reported) — so this verdict is contested.
Highest Terminal-Bench 4.0 resolution rate — the coding test inside the AA index.
03Best value per dollar
Xiaomi (MiMo)
Index 46 at $0.13 per task = 354 points per dollar.
Next: Claude Haiku 5.5 (max) 205, GPT-6.1 Sol (high) 156, Claude Opus 5.5 (medium) 38.
Caveat Among the eleven major labs the pick is Claude Haiku 5.5 (max): index 43 at $0.21 per task (205 points per dollar).
Most AA index points per dollar of AA cost-per-task, among settings scoring 40 or higher.
04Fastest
Meta
245 tokens/sec at index 45, first token after 22.3s.
Next: Claude Haiku 5.5 (max) 237.9, Gemini 3.8 Flash (high) 141.4, Claude Sonnet 5.5 (max) 135.
Caveat Raw speed champion regardless of intelligence: Celeris-1 at 1,522 tokens/sec, index 6.
Highest measured output tokens per second among settings scoring 40 or higher on the AA index.
05Best open-weight
Xiaomi (MiMo)
Index 46, licence: MIT.
Next: GLM-5.3 45 (GLM-5.3 License (commercial use with restrictions)), Kimi K3 44 (Kimi K3 License (commercial use with restrictions)), DeepSeek V4.1 Flash 39 (MIT).
Caveat Open weights means you may host it yourself; check each licence — several allow commercial use only with restrictions.
Highest AA Intelligence Index among models whose weights are published.
06Best for voice / agents
OpenAI
Index 83.2, 74.5% on the agentic voice test, first audio in 1.34s, about $5.83 per hour.
Next: Gemini 3.8 Live Extended Thinking (High) 82.6, Grok Voice Think Fast 2.0 High 81.3. Cheapest in the top six: Gemini 3.8 Live at $0.84 per hour.
Caveat For text agents doing office work, Claude Opus 5.5 leads GDPval-AA at 1,846 Elo, ahead of Claude Sonnet 5.5 at 1,844.
Highest Artificial Analysis Speech-to-Speech Index (reasoning, conversation flow and tool use, measured by voice).
Labs tracked
15
11 major, 4 to watch
Models in focus
29
28 independently scored
Top intelligence score
58
Claude Opus 5.5
Latest model release
12 Oct
Muse Wonder Full
Cheapest capable output / 1M
$0.50
Claude Haiku 5.5 · index 43
29/ 29
02The picture
Flip the switches above and every chart and the leaderboard redraw together. Each mark is a figure from the data file; a model with no verified figure is left out of that chart and named underneath it. Click any model to send it to the compare tray.
Intelligence vs price
Up and to the left is better: smarter for less. Models on the line have nothing both smarter and cheaper. The price axis is logarithmic — each labelled gridline is ten times the last.
Speed
Release timeline
One lane per lab, newest shipper on top. Dates are to scale.
03Frontier model leaderboard
Click a column to sort. Index is the Artificial Analysis Intelligence Index — higher is better. Prices are US dollars per million tokens. A † marks a figure with a caveat; hover or tap it. A dash means no verified figure.
| Claude Opus 5.5max | Anthropic | Closed | 58 | 66.4%† | 1,846† | $4 | $20 | $5.98 | 1M | 95.2 | 691.7s† | 2026-09-22 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude Sonnet 5.5max | Anthropic | Closed | 56 | 70.6%† | 1,844† | $2 | $10 | $5.46 | 1M | 135 | 451.9s | 2026-09-28 |
| Muse Wonder Fullmax | Meta | Closed | 56 | — | — | — | — | — | 1M | — | — | 2026-10-12 |
| Claude Fable 5.1max | Anthropic | Closed | 53 | 55.8%† | 1,735† | $10 | $50 | $7.63 | 1M | 68 | 246.6s | 2026-09-01† |
| GPT-6 Astramax | OpenAI | Closed | 53 | 57.9%† | 1,542† | $10 | $50 | $3.26 | 1.05M | 63 | 326.1s | 2026-09-03† |
| Gemini 4 Argonhigh | Google / DeepMind | Closed | 53 | — | — | $2† | $10† | $1.99 | 1M | — | — | 2026-09-30† |
| GPT-6.1 Solmax | OpenAI | Closed | 52 | — | — | $2 | $10 | $0.72 | 1.05M | 65 | 321.8s | 2026-09-29† |
| Muse Wondermax | Meta | Closed | 52 | — | — | $4† | $20† | $4.52 | 1M† | 101.4 | 19.7s | 2026-10-12 |
| Muse Spark 1.3max | Meta | Closed | 48 | — | — | $1.25† | $4.25† | $1.60 | 1M | 123.9 | 40.8s | 2026-09-02 |
| Grok 4.7xhigh | xAI (SpaceXAI) | Closed | 46 | 37.6%† | 1,695† | $2† | $6† | $3.74 | 500k | 83.8 | 87.0s | 2026-09-21 |
| MiMo-V2.6-Proreasoning | Xiaomi (MiMo) | Open | 46 | — | — | $0.43 | $0.87 | $0.13 | 1M | 47.9 | 3.7s | 2026-09-21† |
| Qwen3.8 Maxreasoning | Alibaba (Qwen) | Closed | 45 | — | — | $2† | $6 | $5.41 | 984k | 40.6 | 2.9s | 2026-09-02† |
| GLM-5.3max | Z AI (GLM) | Open | 45 | — | — | $1.40 | $4.40 | $2.01 | 1M | 66.9 | 3.3s | 2026-08-18† |
| Kimi K3max | Moonshot AI (Kimi) | Open | 44 | — | — | $3† | $15 | $2 | 1.04858M | 43 | 3.1s | 2026-07-16 |
| Claude Haiku 5.5max | Anthropic | Closed | 43 | 39.2%† | 1,620 | $0.10† | $0.50† | $0.21 | 1M | 237.9 | 309.8s† | 2026-10-07 |
| Gemini 3.8 Flashhigh | Google / DeepMind | Closed | 41 | — | — | $0.75† | $3.75† | $1.24 | 1M | 141.4 | 22.7s | 2026-09-02 |
| DeepSeek V4.1 Flashmax | DeepSeek | Open | 39 | — | — | $0.30† | $1.20† | $0.27 | 1M | 197.3 | 1.1s | 2026-09-10 |
| GPT-6 Lunamax | OpenAI | Closed | 38 | 16.4%† | 1,437 | $0.10 | $0.50 | — | 1.05M | 134.2 | 113.3s | 2026-09-22 |
| Mistral Large 4preview | Mistral | Closed | 38 | 28.3%† | — | $1.36 | $4.18 | $1.13 | 524k | — | — | 2026-10-06 |
| DeepSeek V4 Pro 0813max | DeepSeek | Open | 36 | — | — | $1.32† | $3.96† | $0.67 | 1M | 95 | 1.7s | 2026-08-13† |
| Qwen3.8 27Bxhigh | Alibaba (Qwen) | Open | 34 | — | — | $0.50† | $3 | $1.01 | 256k | 45.9 | 3.8s | 2026-08-14† |
| Kimi K2.7 Codedefault | Moonshot AI (Kimi) | — | 26 | — | — | $0.95† | $4 | $0.54 | 262.144k | 75 | 2.8s | — |
| Gemini 3.5 Flash-Litedefault | Google / DeepMind | — | 22 | — | — | $0.30 | $2.50 | $0.19 | 1M | 408 | 7.9s | — |
| Muse Glimmer 30Bhigh | Meta | Open | 17 | — | — | $0.33† | $1.35 | $0.06 | 131k | 133.8 | 1.0s | 2026-08-10 |
| Gemma 4 31Breasoning | Google / DeepMind | Open | 15 | — | — | — | — | — | 256k | 35.5 | 1.0s | 2026-04-02 |
| Nova 2.0 Pro (Preview)medium | Amazon | Closed | 14est† | — | — | $1.38† | $11 | — | 256k | 122 | 13.2s | 2025-11-27 |
| Mercury 2.5reasoning | Inception | Closed | 12 | — | — | $0.25 | $0.75 | $0.12 | 260k | 705.8 | 3.4s | 2026-09-08 |
| Mistral Small 4reasoning | Mistral | — | 11 | — | — | — | — | $0.02 | 256k | 158 | 0.8s | — |
| Celeris-1default | Celeris | — | 6 | — | — | — | — | $0.05 | 131k | 1,522 | 0.6s | — |
The input filter only matches models whose input types we verified; unverified models drop out when you pick one. “est” = Artificial Analysis marks the score as an estimate. 214 of 493 tracked data points are blank for lack of a verifiable source.
04Head to head
Pick models here, from the leaderboard, or by clicking any chart. The radar is relative to your picks: on each spoke the best of the models you chose reaches the rim and the others are drawn as their true share of it.
05For business owners
Pick the job. The answer is computed from the data on this page, so it moves when the leaderboard moves.
06Voice agents
Speech-to-speech models are scored separately. S2S index and the agentic score come from Artificial Analysis; cost is per hour of input audio.
| Voice model | Lab | S2S index | Agentic | First audio | $ / hour |
|---|---|---|---|---|---|
| GPT-Live-1 (Astra, medium) | OpenAI | 83.2 | 74.5% | 1.34s | $5.83 |
| Gemini 3.8 Live Extended Thinking (High) | Google / DeepMind | 82.6 | 68.6% | 1.35s | $3.50 |
| Grok Voice Think Fast 2.0 High | xAI (SpaceXAI) | 81.3 | 56.5% | 0.7s | $4.80 |
| GPT-Live-1 (Sol, low) | OpenAI | 80.1 | 59.3% | 1.24s | $4.47 |
| Gemini 3.8 Live | Google / DeepMind | 76.0 | 30.1% | 1.18s | $0.84 |
| GPT-Realtime-2.1 High | OpenAI | 75.2 | 45.7% | 1.21s | $10.75 |
07The labs
Latest · 7 Oct 2026Shipped Claude Haiku 5.5 at $0.10 / $0.50 per million tokens for prompts up to 100k tokens — a tenth of the previous Haiku's price. source
Latest · 29 Sep 2026GPT-6.1 Sol is listed in the API at $2 / $10 per million tokens, described by OpenAI as near-Astra performance at lower cost. source
Latest · 30 Sep 2026Announced Gemini 4 Argon. Access starts with vetted cyber defenders, then paid API customers and AI Ultra subscribers; introductory price $2 / $10. source
Latest · 12 Oct 2026Released Muse Wonder (Meta Model API and Muse Code) and Muse Wonder Full (Muse Code subscribers only). source
Latest · 21 Sep 2026Released Grok 4.7 at $2 / $6 per million tokens, with a fast variant at twice the speed and twice the price. source
Latest · 6 Oct 2026Put Mistral Large 4 (1 trillion parameters, 52B active) into public preview at $1.36 / $4.18; open weights promised by the end of October 2026. source
Latest · 10 Sep 2026Released DeepSeek V4.1 Flash with weights on Hugging Face; off-peak API rates are half of peak. source
Latest · 16 Jul 2026Kimi K3 (2.8T parameters, 104B active) is the current flagship, with open weights under the Kimi K3 License. source
Latest · 2 Sep 2026Qwen3.8 Max (0902 snapshot) lists at $2 / $6 on Alibaba Cloud's international region. source
Latest · 27 Nov 2025Nova 2.0 Pro remains in preview; Artificial Analysis still marks its intelligence score as an estimate. source
Latest · 1 Oct 2026Launched MAI-Transcribe-2-Streaming ($0.54 per audio hour, introductory) plus MAI-Voice-2.1 ($22 per million characters) and a Flash voice ($15). Microsoft's current first-party push is voice, not a frontier chat model. source
Latest · 18 Aug 2026GLM-5.3 scores 45 on the Artificial Analysis index with open weights. source
Latest · 21 Sep 2026MiMo-V2.6-Pro: MIT-licensed open weights scoring 46 on the Artificial Analysis index at $0.13 per index task. source
Latest · 8 Sep 2026Mercury 2.5 measured at 705.8 output tokens per second. source
Latest · —Celeris-1 measured at 1,522 output tokens per second — the fastest row on the leaderboard, with a low intelligence score. source
08On the radar
Meta
Wonder is on the Meta Model API and Muse Code; Wonder Full is limited to Muse Code subscribers. Artificial Analysis scores Wonder Full at 56, tied for second overall.
Anthropic
Two price tiers split at 100k prompt tokens. Artificial Analysis scores it 43 at 237.9 tokens per second.
Mistral
A 1-trillion-parameter multimodal model at $1.36 / $4.18. Mistral says open weights follow by the end of October.
Microsoft
MAI-Transcribe-2-Streaming covers 60 languages at $0.54 per audio hour (introductory); two new text-to-speech voices at $22 and $15 per million characters.
Introductory $2 / $10, rising to $4 / $20. Google reports 77.9% on DeepSWE v1.1. Rollout starts with vetted cyber defenders; not generally available yet.
OpenAI
Listed at $2 / $10 with a 1.05M-token context. Artificial Analysis scores it 52 at $0.72 per index task.
Anthropic
$2 / $10. Anthropic reports 70.6% on Terminal-Bench 4.0, above its own Opus 5.5.
Anthropic
$4 / $20. Tops the Artificial Analysis Intelligence Index at 58.
OpenAI
Luna is the low-cost tier at $0.10 / $0.50; Artificial Analysis scores it 38.
xAI
$2 / $6 with a 500k-token context. xAI reports 71.0% on DeepSWE v1.1 and 37.6% on Terminal-Bench 4.0.
Industry
After Anthropic's CEO published a 12 September essay urging the industry to pace frontier work — backed the same day by the heads of OpenAI, xAI and Google DeepMind — paying subscribers filed a proposed class action in the Northern District of California alleging illegal coordination. No ruling has been made.
DeepSeek
A 552B-parameter mixture-of-experts model. Artificial Analysis scores it 39 at 197.3 tokens per second and lists the licence as MIT.
OpenAI
OpenAI's top tier at $10 / $50. Artificial Analysis scores it 53.
$0.75 / $3.75 through 31 December 2026, then $1.50 / $7.50 from 1 January 2027.
Anthropic
Fable 5.1 is $10 / $50. Mythos 5.1 is offered only through verified cyber and life-science programs.
Mogul AI
We pick the right model for each job in your business, wire it to your phones, website and follow-up, and keep it current as the leaderboard changes.
09Method
Each “Who's winning” verdict is a single stated rule applied to the table — no editorial weighting. A model counts as “capable” for the value and speed verdicts when it scores 40 or higher on the Artificial Analysis Intelligence Index. Scores marked as estimates are excluded from every verdict and from the price chart.
Independent measurements are preferred over a lab's own claims. Where only the lab has published a number, it is labelled vendor-reported. Where two sources disagree, both are kept in the note on that figure. Release dates taken from a tracker rather than the lab's own page are flagged as such.
Charts are drawn to scale from the same data file as the table. Nothing is interpolated: a model with no verified figure for a chart is left out of it and listed in the line beneath. The Reasoning switch set to Low swaps in the lowest reasoning setting Artificial Analysis measured, for the models where one was measured; the rest stay at their only measured setting.
10Sources
Train Smarter. Perform Stronger. Recover Faster.
Wireless EMS training suit — 20-minute full-body sessions, app included, no monthly subscription.
The Adapt EMS Dry Suit — a black wireless EMS training suit with the Adapt EMS Suits logo on the chest.
See the suits adaptemssuits.com
MOGUL
