Lewis - Policifyai (@Policifyai) on X

What it claims

This is an October 2026 month roundup posted on X by Lewis (@Policifyai). It cites no primary sources. Its main model claims are four. GPT-6.1 Sol beats GPT-6 Astra on several agentic coding benchmarks, including DeepSWE, at lower cost. Claude Sonnet 5.5 outscores Claude Opus 5.5 on Terminal-Bench 4.0 in some configurations. Gemini 4 Carbon leads Artificial Analysis Intelligence Index v5 but has restricted access. Mistral Large 4 ("Le Chonk") is the strongest open-weight model outside China. The author attributes OpenAI's repeated price cuts to pressure from open Chinese models.

The Anthropic S-1 figures the summary does not give are these. Anthropic would list on Nasdaq under the ticker ANT. It reported $16.3 billion in first-half 2026 revenue against a net loss of about $80.7 billion, mostly non-cash. It has up to $120.5 billion in NVIDIA GPU capacity from SpaceX.AI LLC through 2029, all expected by December 2026. It also has a minimum $110 billion Amazon commitment running from May 2026 to April 2036. Founders keep majority voting power, and the Long-Term Benefit Trust controls a majority of board seats.

Evidence

On Terminal-Bench 4.0, Sonnet 5.5 scored 70.6% and Opus 5.5 66.4%, "in some configurations." GPT-6.1 Terra scored 50.3% at maximum reasoning effort, up from 21.5% for the previous Terra. Le Chonk reached 41%, a 13-point gain the page credits to post-preview reinforcement learning. Haiku 5.5 scored 39.2% on Terminal-Bench 4.0 and 72.4% on OSWorld.

Intelligence Index v5 uses nine evaluations across four areas. 60% of the score comes from private or held-out tests, and new benchmarks include FrontierCode 2.0 and Terminal-Bench-Science. Initial scores were Opus 5.5 (max) 58, GPT-6 Astra (max) 57 and GPT-6.1 Sol (max) 55. Gemini 4 Carbon later scored 59. The best open-weight model trails by about 12 points. Scores are not comparable with earlier index versions. Le Chonk leads open-weight models on FrontierCode 2.0 at 19.7%.

Gemini 4 Carbon posts 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. Artificial Analysis measured Vera Rubin NVL72 at up to 5.7 times the throughput per megawatt of GB300 NVL72 on DeepSeek V4 Pro. The lead narrows, but stays positive, at peak throughput. Nemotron 3.5 Transcribe has a word error rate around 2.8% across 40 languages.

How to use it

Closed-model prices are per million input/output tokens:

• GPT-6.1 Sol: $2/$10, with cached input at $0.10.
• Sonnet 5.5: $2/$10, with a free tier.
• Haiku 5.5: $0.10/$0.50.
• Opus 5.5: $4/$20, with 1 million tokens of context.
• Gemini 4 Argon: $2/$10 on a promo rate ending around the end of November 2026, with 1.05 million tokens of context. It is generally available on the Gemini API, AI Studio, the Gemini app for Pro and Ultra users, and Antigravity.
• Grok 4.7: $2/$6, with 500k context.

Grok Voice 3.0 costs $0.08 per minute.

GPT-6.1 Terra uses the API ID gpt-6.1-terra. It has a 1,050,000-token context window, takes text and image input, and has a knowledge cutoff of 1 June 2026. It is available in the OpenAI API, on Amazon Bedrock, and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. OpenAI still recommends Sol and Astra for the hardest coding tasks. GPT-6 Astra, Sol and Luna are in private preview on Google Cloud Model Garden for select enterprise customers. OpenAI's Decisions API is in public beta for fixed-choice answers. UltraFast mode on Cerebras serves Astra at about 300 tokens per second on the Pro tier.

The open-weight releases the page sizes are these:

• Le Chonk: about 1 trillion total parameters and 50 billion active, 1 million context, multimodal. It is on Hugging Face and the Mistral API, and the preview cost about $1.13 per task in some reports.
• Muse Spark Open: 816B total, 42B active, 1 million context, multimodal input with text output.
• Kimi K3.1 Pro: 2.8T total, 104B active, 1 million context. Weights are promised for early November 2026.
• Aleph Alpha Kolibri-1: 78B total, 3.46B active, German-English, 1 million context, Apache 2.0.
• Nemotron 3.5 Super VL: 120B total, 12B active.
• Liquid AI d1-3B and d1-omni-600M: decisions in 8 ms on a GPU and about 50 ms on a Jetson.

OpenAI, Google, Microsoft and Anthropic back a shared robots.txt signal that lets sites opt out of foundational generative AI training.

Limits

The post is a secondary roundup. Most benchmark settings are unstated or hedged with phrases such as "in some configurations" or "some reports." Dates are approximate. Kimi K3.1 Pro ("around the 23rd") and OpenAI on Google Cloud ("around the 21st") are filed under mid-October (11 to 20).

Several releases carry stated restrictions:

• Gemini 4 Carbon is limited to cyber defenders in Google's Fairwind Program, with no date for broad access.
• Muse Spark Open carries a revenue-cap licence.
• Le Chonk showed potential cybersecurity uplift in Mistral's internal monitoring, and Mistral is tightening guardrails.
• Terra meets the Critical cybersecurity threshold in OpenAI's Preparedness Framework.
• Dots, OpenAI's always-on agent system, is limited to the Pro tier.
• Genie 4 is open only to trusted testers. Sessions last a few minutes, object permanence drifts, and real places are not reproduced accurately.

A GPT-6 Luna update raised factual-error rates on some internal measures.