2026-09-23

1. Anthropic shipped a model whose whole pitch is costing less — and dropped SWE-bench from the launch table doing it (1.69)

• Claude Opus 5.5, released 2026-09-22 as claude-opus-5-5: $4/M input · $20/M output, cache reads $0.20/M, 1M context, 128K max output, adaptive thinking always on at a default effort of medium, training cutoff Jun 2026. Day-0 on the Claude apps, Claude Code, the API, Bedrock, Vertex and Microsoft Foundry (source)
• Anthropic's claim is Fable 5.1-level work at ~40% lower cost than Opus 5, and the stated mechanism is fewer tokens per task and output 30%+ faster, not the per-token price alone. The docs now say start here for most workloads
• Launch table: Terminal-Bench 4.0 66.4% (Fable 5.1 55.8%, Opus 5 52.3%), Terminal-Bench-Science 0.1 58.7% against Opus 5's 29.0%, Humanity's Last Exam 67.7%, GDPval-AA v2.1 1846 Elo, AutomationBench 40.0%
• Ships preserved thinking, an anti-distillation safeguard that stops API users editing Claude's prior context to extract its reasoning — a capability restriction shipped as a headline feature
• Why it matters: a year of entries on Anthropic measured each release against the last on SWE-bench. This table has no SWE-bench-family row at all, which makes it uncomparable to the DeepSWE figure OpenAI published the same day — the benchmark divergence is now the story, not a footnote
• Not established: no effort level is stated for any row, and this model defaults to medium where Fable 5.1 defaults to high; the 40% and 30% figures are Anthropic's own and unmeasured here
• → Claude Opus 5.5 · Anthropic · Eval Harness Configuration

2. OpenAI halved GPT-6 prices twice over, and published exactly one benchmark to justify it (1.63)

• GPT-6 Sol and GPT-6 Luna, released 2026-09-22, 19 days after GPT-6 Astra. Both: 1.05M context, 128,000 max output, reasoning none through max (source)
• Sol $2/$10 against GPT-5.6 Sol's $4/$20. Luna $0.10/$0.50 against GPT-5.6 Luna's $0.20/$1.20. Cached input takes a 90% discount — $0.20/M and $0.01/M
• DeepSWE v1.1 is the entire published record: Sol 68.8% at max effort, Luna 66.6% at max, against Claude Fable 5's 69.9% at xhigh. OpenAI puts Sol's cost per task ~80% below Fable 5's, and Luna's at 93% below Opus 5 and 96% below Fable 5
• Free and Go users get Luna in the desktop app — the first GPT-6 model to reach unpaid tiers. Neither model is in Chat yet
• Why it matters: every Astra entry on OpenAI has been about a capability ceiling and the Preparedness Framework treatment that delayed it. This is the inverse — ceiling untouched, floor halved, cheapest tier given away
• Not established: one benchmark means nothing can be said about either model outside software engineering; Sol's max and Fable 5's xhigh are each vendor's top setting but not the same setting; and nothing read states how either model is classified under the Preparedness Framework — a real gap 19 days behind the first model OpenAI confirmed at the Critical cybersecurity threshold
• → GPT-6 Sol · GPT-6 Luna · OpenAI

3. The top open-weights model in the world now belongs to a phone manufacturer — with its training environments attached (1.40)

• Xiaomi released MiMo-V2.6-Pro on 2026-09-22: 1.02T total / 42B active sparse MoE, 1M context, omnimodal (text, image, video, audio in), MIT on Hugging Face — with MiMo-V2.6-Flash (309B/15B) and a 9B distillation (source)
• The RL stack and training environments were open-sourced too. Publishing weights is routine on this wiki; publishing the reinforcement-learning environments that produced them has no precedent here
• Reported 46 on Artificial Analysis' Intelligence Index v4.3.2, first among open-weights models, ahead of GLM-5.3 (45) and Kimi K3 (44). Latent Space headlines the Pro checkpoint as "trained for $3M"
• Why it matters: the open-weights lead has passed between DeepSeek, Z.ai and Moonshot AI all year and has just gone to a company none of those pages mention — under MIT, at a claimed training cost two orders of magnitude below any frontier figure this wiki holds
• Not established, and it matters here: the 46 is third-party reporting of an Artificial Analysis page, not this repo's own Sunday snapshot, which predates the release — the 2026-09-27 scrape is what would confirm it. A one-point lead on a 10-evaluation composite is inside the range where ordering is not a finding. No vendor benchmark table was obtained at all, and the $3M figure is a characterisation, not a Xiaomi statement
• → MiMo-V2.6-Pro · Xiaomi · Open-Weights Policy Fight