Quick answer

September 2026 was the busiest month for frontier AI releases so far. In order: Claude Fable 5.1 and Mythos 5.1 (Sept 1), Gemini 3.8 Flash (Sept 2), Meta's Muse Spark 1.3 (Sept 2), GPT-6 Astra and GPT-5.7 (Sept 3), Meta Muse (Sept 8), Astra for Law (Sept 17), Qwen3.8-Omni-Flash (Sept 18), Grok 4.7 (Sept 21), and Claude Opus 5.5 (Sept 22). The common thread is price: almost every launch came with a cut, an introductory rate, or a cheaper tier, and the frontier is now available from $2 to $10 per million input tokens depending on the vendor.

If you only read one thing about AI this month, read this: the models got a bit better and a lot cheaper. Three years ago a frontier model cost $30 to $60 per million output tokens. This month Grok 4.7 shipped at $6, Opus 5.5 at $20, Gemini 3.8 Flash at $3.75, and Qwen3.8-Omni-Flash at $0.47. Capability differences between the top models are now measured in single-digit benchmark points, while price differences are measured in multiples. That inversion is the story of September.

The launches, in order

  • Sept 1 — Claude Fable 5.1 and Mythos 5.1 (Anthropic): the same model at two safeguard levels, Mythos gated to trusted-access programmes. Pricing unchanged at $10/$50 per million tokens; cache reads cut 75% to $0.25. 1M input / 128K output context
  • Sept 1 — Perplexity Hybrid Compute on Mac: Pro, Max, and Enterprise users get a local 27B model that handles part of each task on-device, routing the rest to the cloud
  • Sept 2 — Gemini 3.8 Flash (Google): the third Flash variant in six weeks, at an introductory $0.75/$3.75 until December 31, rising to $1.50/$7.50 in January. A gated Flash Cyber variant for governments and infrastructure operators
  • Sept 2 — Muse Spark 1.3 (Meta): developer model at $1.25/$4.25 with a $0.10/$0.20 "Contributor" tier for those who share training data. About 20% fewer tool calls and 25% fewer tokens than 1.2
  • Sept 3 — GPT-6 Astra and GPT-5.7 (OpenAI): Astra at $10/$50 with computer use and the first "Critical" cybersecurity rating in OpenAI's framework; GPT-5.7 as the broadly available upgrade at unchanged $5/$15
  • Sept 8 — Meta Muse: a personal agent that runs errands and fills forms in a secure VM, free up to 100 million tokens a week, with $20 and $100 tiers. US-only, 18+
  • Sept 17 — Astra for Law (OpenAI): GPT-6 Astra with a legal search index, lifting correctness on legal questions from 38.7% to 54.0% in OpenAI's own evaluation
  • Sept 18 — Qwen3.8-Omni-Flash (Alibaba): a native omnimodal model taking text, image, audio, and video input at $0.15/$0.47 per million tokens
  • Sept 21 — Grok 4.7 (xAI): a larger base model with extended reinforcement learning, a 500K context, and pricing of $2/$6 — the cheapest frontier-tier launch of the month
  • Sept 22 — Claude Opus 5.5 (Anthropic): Fable 5.1-level performance on most work at $4/$20, 40% cheaper to run than Opus 5, with thinking that can no longer be switched off

What the month means if you use AI

  • Consumers: the assistant you already pay for got better without a price rise. ChatGPT, Claude, and Gemini subscribers all received new default models this month
  • Agents are now a product category, not a demo: Meta Muse, GPT-6 Astra's computer use, and Opus 5.5's agentic-coding lead all ship to the public
  • Local AI went mainstream by stealth: Perplexity routing work to a model on your Mac is the first time a major consumer product made hybrid local-cloud compute the default
  • Vertical models arrived: Astra for Law is the first frontier model configured for a profession by its maker rather than a startup on top

What it means if you build with AI

Re-run your cost model. If your product was designed around 2025 prices, you can probably afford a stronger model tier today for the same budget, or the same tier for a fraction of the cost. Check the fine print on the cheapest offers: Gemini's price doubles in January, Meta's Contributor tier trades data for discount, and Opus 5.5 charges for thinking on every request. And watch the breaking changes — Opus 5.5 rejects several older API parameters, and every vendor is pushing developers toward adaptive reasoning controls and structured outputs.

The benchmark race is real but increasingly irrelevant to most decisions. When Terminal-Bench scores differ by nine points and prices differ by five times, price and fit win. Pick by job, then by cost, then by benchmark.

What did not happen

  • No new open-weight frontier model from Meta — Muse Spark is API-only, continuing the shift away from open Llama releases
  • No price rise from anyone, despite record compute demand — every move was down or flat
  • No independent benchmark confirmations yet for the September models; every number above is vendor-reported

Related reading

Claude Opus 5.5 Released — Fable-Level Work at 40% Less GPT-6 Astra Released — OpenAI's Shift From Answering to Doing Meta Muse Launched — A Personal AI Agent That Runs Your Errands Grok 4.7 Released — xAI Undercuts Everyone at $2 per Million AI API Pricing in 2026 — The Honest Comparison

Bottom line

September 2026 compressed a year of releases into three weeks and moved the frontier down-market. The practical advice is unglamorous: revisit which model you use and what you pay for it, because the answer from June is probably wrong now — and Sonnet 5.5 and Haiku 5.5 are due within weeks to move it again.