AI Model Comparison

Frontier and open-weight models compared on context, price, scale, benchmarks and risk.

Updated August 2026

Model Provider Context / max out Price in / out per 1M Parameters Training compute Benchmark Capabilities Resource Operational risk Provider share · MAU
Claude Fable 5 Anthropic 1M / 128K $10 / $50 Not published Not published 95.0% SWE-bench V · SOTA ⁸ 8 / 8 API only 30-day retention required; refusals 9.2% · 245M
Claude Opus 5 Anthropic 1M / 128K $5 / $25 Not published Not published — 8 / 8 API only Refusals return HTTP 200 9.2% · 245M
Claude Sonnet 5 Anthropic 1M / 128K $2 / $10 Not published Not published — 8 / 8 API only Tokenizer counts ~30% more 9.2% · 245M
Claude Haiku 4.5 Anthropic 200K / 64K $1 / $5 Not published Not published — 6 / 8 API only 200K context ceiling 9.2% · 245M
GPT-5.6 Sol OpenAI 1.05M $5 / $30 ¹ Not published Not published 94.1% GPQA · SOTA (tied) Multimodal API only No cut while tiers below fell 53.9% ² · 1.11B
GPT-5.6 Terra OpenAI 1.05M $2.50 / $15 Not published Not published — Multimodal API only Cut 20% on 30 Jul 2026 53.9% ² · 1.11B
GPT-5.6 Luna OpenAI 1.05M $1 / $6 Not published Not published — Multimodal API only Cut 80% on 30 Jul 2026 53.9% ² · 1.11B
GPT-5.5 OpenAI — $5 / $30 Not published Not published 93.5% GPQA Multimodal API only Sol's price, lower score 53.9% ² · 1.11B
GPT-5.4 OpenAI — $2.50 / $15 Not published Not published 75.0% OSWorld-V ⁸ Multimodal API only Deprecation clock 53.9% ² · 1.11B
GPT-5.4 Nano OpenAI — $0.20 / $1.25 Not published Not published — Text-first API only Low quality ceiling 53.9% ² · 1.11B
Gemini 3.1 Pro Google 1M $2 / $12 ³ Not published Not published 94.1% GPQA · SOTA (tied) Multimodal API only Preview; price doubles >200K 27.9% · 662M
Gemini 3.7 Flash Google 1M $0.75 / $3.75 ⁴ Not published Not published — Multimodal API only Promo rate doubles Dec 2026 27.9% · 662M
Gemini 3.6 Flash Google 1M $1.50 / $7.50 ⁵ Not published Not published — Multimodal API only None material 27.9% · 662M
Gemini 3.5 Flash-Lite Google 1M $0.30 / $2.50 Not published Not published — Multimodal API only None material 27.9% · 662M
DeepSeek V4 Pro Open weight 1M Weights free 1.6T / 49B active Not published 80.6% SWE-bench V ⁸ Text + reasoning Very high You own serving + safety stack —
GLM-5.2 Open weight — Weights free — Not published 81.0 Terminal-Bench ⁸ Text + code High Check licence terms —
GLM-5.1 Open weight — Weights free 744B / 40B active Not published Opus-class coding Text + code High Trained on Huawei Ascend —
Qwen 3.7 Open weight 128K ⁶ Weights free 397B sparse MoE ⁷ Not published GPT-5.5-class reasoning Text · VL variants High Param figure from 3.5 sibling —
Qwen 3.6 Open weight — Weights free — Not published — Text Moderate Not a frontier contender —
Llama 4 Maverick Open weight — Weights free 400B / 17B active Not published — Text Moderate None material —
Llama 3.3 70B Open weight — Weights free 70B dense Not published — Text Low Older generation —

SOTA* = the highest value in this table* for that specific benchmark as of the update date, not an absolute claim; "open SOTA" leads among open-weight models only. Scores come from different dates and harnesses, so they are not a controlled head-to-head · "Not published" = the lab has never disclosed it, true of every parameter count and training-compute figure for every closed model · "—" = did not surface in the sources below · Capabilities scored against eight items: text, image input, ≥200K context, tool calling, structured outputs, extended reasoning, hosted code execution, hosted web search; only the Anthropic column was checked against first-party docs · Share and MAU are per provider, not per model (Similarweb, May 2026) · ¹ $0.50 cached input · ² a later mid-2026 reading puts this at 46.4% · ³ up to 200K context; $4 / $18 above · ⁴ promotional through Dec 2026, then $1.50 / $7.50 · ⁵ $0.15 cached input · ⁶ 256K with YaRN on the 397B-A17B variant · ⁷ figure carried over from the 3.5 generation · ⁸ contested or partial — see the per-model benchmark table below · Frontier 2026 training runs are estimated at 1e26–1e27 FLOP and $200M–$500M per run; GPT-4 was roughly 2e25 in 2023

Benchmark scores by model

Model Licence TB 2.0 TB 2.1 OSWorld-V ARC-AGI-2 HLE GPQA SWE-bench V
Claude Mythos 5 Closed 88.0% — 85.0% — 64.5% — —
Claude Fable 5 Closed 84.3% — 85.0% — 64.5% — 95.0% ⁹
Claude Opus 5 Closed — 89.1% — — — — 97.00%
Claude Sonnet 5 Closed 80.4% — 81.2% — 57.4% — —
Claude Opus 4.8 Closed 74.6% — 83.4% — 57.9% — —
Claude Opus 4.7 (Adaptive) Closed — — 78.0% 75.8% 54.7% — —
Claude Opus 4.6 Closed — — — — 53.0% 91.3% —
GPT-5.6 Sol (xhigh) Closed — 89.5% — — — 94.1% —
GPT-5.5 Closed 82.0% — 78.7% 85.0% — 93.5% —
GPT-5.5 Pro Closed — — — — 57.2% — —
GPT-5.4 Closed — — 75.0% — — — —
GPT-5.4 Pro Closed — — — 83.3% 58.7% — —
GPT-5.2 Closed — — — 52.9% — — —
Gemini 3.1 Pro Closed — — — 77.1% — 94.1% —
Gemini 3.7 Flash Closed — 85.8% — — — — —
Gemini 3.5 Flash Closed 76.2% — 78.4% 72.1% — — —
Gemini 3 Pro Deep Think Closed — — — 45.1% — — —
Gemini 3 Pro Closed — — — 31.1% — — —
Grok 4.6 (high) Closed — 88.4% — — — — —
Grok 4.20 Closed — — — 53.3% — — —
Sakana Fugu-Ultra Closed 82.1% — — — — — —
Sakana Fugu Closed 80.2% — — — — — —
Holo3-122B-A10B Closed — — 78.8% — — — —
Muse Spark Closed — — — 42.5% — — —
DeepSeek V4 Pro 0813 Open — 87.9% — — — — 96.40% ⁹
Qwen3.8 Max Open — 86.6% — — — — —
GLM-5.2 Open 81.0% — — — 54.7% — —
GLM-5.1 Open — — — — 52.3% — —
Holo3-35B-A3B Open — — 82.6% — — — —
Ornith-1.0-397B Open 77.5% — — — — — —
Kimi K3 Open — — — — — — 93.40%

Bold* = highest retrieved score in that column · "—" = the model does not appear in the slice of that leaderboard retrieved here, which is not a zero and not an absence of capability — most of these leaderboards publish only a top-10 view, and the SWE-bench Verified column is the thinnest of all — its full ranking was behind a bot check, so only the widely-reported top scores are shown · TB 2.0 and TB 2.1 are different benchmarks*, which is why GLM-5.2 leads open weights on one and DeepSeek V4 Pro 0813 leads on the other · ⁹ SWE-bench Verified is contested: a second source puts Claude Fable 5 first at 95.0% and DeepSeek V4 Pro at 80.6%, a 16-point spread most likely explained by the dated 0813 snapshot versus the base model · Scores are configuration-dependent — effort setting and model snapshot are part of the result, so they are named in the Model column where the source gave them · MMLU is omitted: saturated above 88% across the frontier, it no longer differentiates · BenchLM's weighted composite category indices are also omitted — they are not raw scores, and its published top-3 lists are internally inconsistent · Sources: awesome-llm-bench mirror of BenchLM, synced 16 Aug 2026 (TB 2.0, OSWorld-V, ARC-AGI-2 and HLE tables updated 7 Jul 2026); Artificial Analysis for TB 2.1, 14 Aug 2026; llm-stats for SWE-bench Verified

Sources

Anthropic

OpenAI

Google

Benchmarks

Open weights

Traffic and adoption

Training compute