Frontier Log

China vs the Rest

Open-weights flagships by date, sized by total parameters, with GPQA Diamond: Llama 3.1 405B (Meta, 405B), DeepSeek-V3 (671B, 59.1%), Kimi K2 (1T, 75.1%), gpt-oss-120b (OpenAI, 117B, 80.1%), DeepSeek V4-Pro (1.6T, 90.1%) and Qwen3.8-Max (2.4T, 92.6%)

Download

SWE-bench Verified against output price per 1M tokens, with the Pareto frontier: DeepSeek-V3.2-Exp (67.8%, $0.42), MiniMax M3 (80.5%, $2.40), DeepSeek V4-Pro (80.6%, $3.48) and Claude Opus 4.7 (87.6%, $25) on it; GPT-5 (74.9%, $10), GPT-5.2 Thinking (80.0%, $14) and Claude Opus 4.6 (80.8%, $25) below it

Download

SWE-bench Verified by release date, closed models against Chinese open models: the open side first passed 75% four months after the closed side (Claude Sonnet 4.5 to Kimi K2.5) and 80% two and a half months after (Claude Opus 4.5 to MiniMax M2.5)

Download

Steps

EventStep
DeepSeek-V3DeepSeek-V3671B: largest open model yetAnnouncement
Kimi K2Kimi K2First 1T open modelAnnouncement
DeepSeek-V3.2-ExpDeepSeek-V3.2-ExpAPI price cut by over halfAnnouncement
DeepSeek-V3.2DeepSeek-V3.273.1% on SWE-bench VerifiedAnnouncement
Kimi K2.5Kimi K2.576.8% on SWE-bench VerifiedAnnouncement
MiniMax M2.5MiniMax M2.580.2% on SWE-bench VerifiedAnnouncement
DeepSeek V4 PreviewDeepSeek V41.6T at $3.48 per 1MAnnouncement
Qwen3.8-MaxQwen3.8-Max2.4T open weightsAnnouncement

Open-weights flagships, total parameters

ModelValueSources
Llama 3.1 405B405BAnnouncementopening: "we're publicly releasing Meta Llama 3.1 405B"
DeepSeek-V3671BAnnouncementWhat's new in V3: "671B MoE parameters"
Kimi K21000BAnnouncementopening: "1 trillion total parameters"
gpt-oss-120b and gpt-oss-20b117BAnnouncementPre-training & model architecture: gpt-oss-120b, "117b" total parameters
DeepSeek V4 Preview1600BAnnouncement"DeepSeek-V4-Pro: 1.6T total / 49B active params"
Qwen3.8-Max2400BAnnouncementmodel summary: "2.4T parameters (95B active), with open weights releasing next week"; Alizila, Aug. 17, 2026: weights released

GPQA Diamond

ModelScoreSources
DeepSeek-V359.1%Announcementbenchmark table image, row 'GPQA-Diamond (Pass@1)', column 'DeepSeek-V3'
Kimi K275.1%Announcementbenchmark table, row 'GPQA-Diamond | Avg@8', column 'Kimi K2 Instruct'
gpt-oss-120b and gpt-oss-20b80.1%Announcementchart 'GPQA Diamond (without tools)', bar 'gpt-oss-120b'
DeepSeek V4 Preview90.1%Announcementbenchmark table image, row 'GPQA Diamond (Pass@1)', column 'DS-V4-Pro Max'
Qwen3.8-Max92.6%Announcement'Full Benchmark Table', row 'GPQA Diamond', column 'Qwen3.8-Max'

Output price, $ per 1M tokens

ModelValueSources
GPT-510Announcement"GPT-5 is priced at $1.25/1M input tokens and $10/1M output tokens"
DeepSeek-V3.2-Exp0.42Announcementprice image: Output $0.42 per 1M tokens
GPT-5.214Announcement"GPT-5.2 is priced at $1.75/1M input tokens and $14/1M output tokens"
Claude Opus 4.625Announcement"Pricing remains the same at $5/$25 per million tokens"
Claude Opus 4.725Announcement"$5 per million input tokens and $25 per million output tokens"
DeepSeek V4 Preview3.48Announcementprice table image: deepseek-v4-pro, Output "$ 3.48" (no unit printed; DeepSeek quotes its API prices per 1M tokens)
MiniMax M32.4Announcementprice table image: MiniMax M3 (≤512K), Output $2.40 per 1M tokens

SWE-bench Verified

ModelScoreSources
Kimi K265.8%Announcementbenchmark table, row 'SWE-bench Verified (Agentic Coding) | Single Attempt (Acc)', column 'Kimi K2 Instruct'
GPT-574.9%Announcementtext, opening summary of results
Claude Sonnet 4.577.2%Announcementbenchmark table (image), row 'Agentic coding, SWE-bench Verified', column 'Claude Sonnet 4.5', first value
DeepSeek-V3.2-Exp67.8%Announcementbenchmark table image, row 'SWE Verified', column 'DeepSeek-V3.2-Exp'
Claude Opus 4.580.9%Announcementbenchmark table (image), row 'Agentic coding, SWE-bench Verified', column 'Opus 4.5'
DeepSeek-V3.273.1%Announcementbenchmark chart image, 'SWE Verified (Resolved)', bar 'DeepSeek-V3.2-Thinking'
GPT-5.280.0%Announcementcoding table, row 'SWE-bench Verified', column 'GPT-5.2 Thinking'
Kimi K2.576.8%Announcementbenchmark chart 'Coding', 'SWE-Bench Verified', bar 'Kimi K2.5'
Claude Opus 4.680.8%Announcementbenchmark table (image), row 'Agentic coding, SWE-bench Verified', column 'Opus 4.6'
MiniMax M2.580.2%Announcementopening paragraph
Claude Opus 4.787.6%Announcementbenchmark table (image), row 'Agentic coding, SWE-bench Verified', column 'Opus 4.7'
DeepSeek V4 Preview80.6%Announcementbenchmark table image, row 'SWE Verified (Resolved)', column 'DS-V4-Pro Max'
MiniMax M380.5%Announcementbenchmark table image, row 'SWE-Bench Verified', column 'MiniMax M3'

Tags