Best AI Models in September 2026

Choose the leader for the job you actually need rather than the largest composite score.

• Hardest generally available coding and knowledge work: Claude Fable 5.1, if its $10/$50 base rate and higher latency are justified.
• OpenAI end-to-end agents and computer use: GPT-6 Astra, once the target account receives rollout access.
• Current high-end daily work: compare Claude Opus 5.5 and GPT-6 Sol using their newer independent evidence and the same acceptance checks; test Grok 4.7 separately because it is outside the selected table.
• High-volume focused work: test GPT-6 Luna where its $0.10/$0.50 base rate and OpenAI tool surface fit, then escalate when the task misses.
• Frontier value: Muse Spark 1.3 xhigh where Meta access is workable, or Gemini 3.8 Flash for a stable GA contract and broad multimodal inputs.
• Restricted variants are not default choices: Muse Spark 1.3 max is limited preview and Gemini 3.8 Flash Cyber requires Fairwind approval.

In Artificial Analysis results current on September 3, Claude Fable 5.1 max scores 66 and Claude Opus 5 max scores 63. Muse Spark 1.3 max scores 62 but is limited preview; its available xhigh configuration scores 61. GPT-6 Astra xhigh and Grok 4.6 high also score 61, while Gemini 3.8 Flash high scores 59. A different reasoning effort, fallback policy, agent harness, or test revision can change the order.

• Fable 5.1: highest selected composite score, high price, and slower output in the same independent comparison.
• Astra: major coding-agent gains, but its broader intelligence-versus-cost result is constrained by $10/$50 token pricing.
• Muse Spark 1.3: strong xhigh value; the higher-scoring max variant is not broadly available and lacks public pricing.
• Grok 4.6: 61 at high reasoning with $2/$6 base pricing below the long-prompt threshold.
• Gemini 3.8 Flash: 59 at high reasoning and a low measured cost per task under the launch promotion.

OpenAI launched GPT-6 Sol and Luna on September 22, Anthropic launched Claude Opus 5.5 the same day, and SpaceXAI launched Grok 4.7 on September 21. Artificial Analysis now compares Opus 5.5 and Sol in its newer Index v4.3.2: max-effort configurations score 58 and 48 respectively, with different reasoning and fallback settings. That new result does not update the older multi-model table or prove a universal coding winner. See the focused Opus 5.5 versus GPT-6 Sol comparison for the exact tradeoffs.

Current base input/output rates per million tokens include $10/$50 for Fable 5.1 and Astra, $4/$20 for Opus 5.5, $2/$10 for GPT-6 Sol, $0.10/$0.50 for GPT-6 Luna, and $2/$6 for Grok 4.7 below 200K prompt tokens. Muse Spark 1.3 xhigh and Gemini 3.8 Flash retain their separately documented rates. Total task cost also depends on threshold pricing, output length, reasoning tokens, tool calls, cache hits, retries, and human correction.

Fable 5.1, Opus 5.5, GPT-6 Sol, GPT-6 Luna, Grok 4.7, and Gemini 3.8 Flash have documented developer contracts. Product access still differs from API access: OpenAI separates Chat, ChatGPT Work, Codex, desktop and API rollout; SpaceXAI separates public Grok 4.7 from product-only Fast access. A model your actual account cannot call is not the best production choice today.

Gemini 3.8 Flash has the broadest documented public input set here: text, image, video, audio, and PDF. Astra, Sol, Luna, Fable 5.1, Opus 5.5, and Grok 4.7 accept text and images with text output. Muse Spark 1.3 is reported with text, image, and video input. Tool ecosystems differ materially: OpenAI emphasizes hosted agent tools and computer use, Google combines grounding and multimodal tools, SpaceXAI includes its own search and product surfaces, and Anthropic emphasizes long-horizon coding and knowledge work.

Start with the cheapest model that meets the task's acceptance checks, then escalate only when a stronger model changes completion quality or review cost. Freeze prompts, tools, permissions, reasoning effort, and validators for the comparison. Record failed attempts and human correction time; otherwise a cheap token rate or a high benchmark score can hide a more expensive production path.