Claude, ChatGPT, Grok, and Muse: Every Major AI Model That Shipped This September

Four of the biggest names in AI shipped new flagship models inside the same three-week stretch this September — Anthropic, OpenAI, xAI, and Meta all had a release with each other’s launches still fresh in the news cycle. That kind of clustering doesn’t happen by accident; it’s a sign of how tight the competitive cycle between these labs has gotten. If you’ve been half-following the headlines and lost track of which model does what, here’s each release in plain language, with a link to the actual announcement for anyone who wants the full detail.

Claude Opus 5.5 (Anthropic) — September 22

Anthropic’s newest model leads with a cost story as much as a capability one: Opus 5.5 performs at roughly the level of the company’s own Claude Fable 5.1 on most everyday work, while costing 40% less to run than the previous Opus 5, according to Anthropic’s own announcement. It’s built for long-running agentic coding and knowledge work specifically, with a 1 million token context window and up to 128k tokens of output in a single response — useful for a task like reviewing an entire codebase or a long research document in one pass instead of splitting it into chunks. Anthropic also reports that Opus 5.5 scored best of any Claude model to date on its internal automated behavioral audit, the alignment testing suite the company runs before every release. Sonnet 5.5 and Haiku 5.5 are expected to follow within weeks, following Anthropic’s usual pattern of staggering a model family’s releases rather than shipping all three sizes at once.

GPT-6 Astra (OpenAI) — September 3

OpenAI’s Astra release carries a genuinely unusual distinction: it’s the first model the company has ever classified “Critical” under its own Preparedness Framework, specifically for cybersecurity risk. Our full breakdown of GPT-6 Astra covers that safety story in depth, but the short version is that Astra is state-of-the-art on computer use, software engineering, cybersecurity, and science, according to OpenAI’s own announcement — strong enough on math benchmarks that OpenAI says it’s already contributed to solving previously open problems. Later in the same month, OpenAI followed up with GPT-6 Sol and Luna on September 22, two models built to bring frontier-level intelligence to everyday work at different balances of capability and cost, rather than a single one-size-fits-all flagship.

Grok 4.7 (xAI) — September 21

xAI’s release leans hardest into raw scale: Grok 4.7 ships on a base model roughly 40% larger than Grok 4.6’s, with a 500,000-token context window and a knowledge cutoff of May 2026, according to xAI’s own announcement. It’s positioned specifically for coding and agentic work that runs for hours rather than a single exchange, with what xAI describes as more careful self-checking along the way. Independent coding benchmarks put it ahead of OpenAI’s smaller GPT-5.6 Sol Max variant but still behind both Claude Fable 5.1 and GPT-6 Astra on the same tests — a useful reminder that a bigger base model and more parameters don’t automatically translate into a benchmark lead over the field’s actual frontrunners, even when the raw numbers on paper look larger.

Muse (Meta) — September 2 and September 8

Meta’s September was arguably the busiest of the four, and it’s worth separating two genuinely different announcements that both use the Muse name. First, Meta Superintelligence Labs shipped Muse Spark 1.3 on September 2 — an incremental but real upgrade to its underlying language model, using roughly 20% fewer tool calls and 25% fewer tokens than the prior version on comparable tasks, according to Meta AI Research’s own announcement. Then, at Meta Connect on September 8, Meta introduced Muse as a consumer-facing personal AI agent product built on top of that model family — a genuinely different thing from the underlying LLM, aimed at everyday users rather than developers, according to Meta’s own announcement. CEO Mark Zuckerberg has described the ambition for Muse as reaching billions of everyday users, which puts Meta’s approach closer to OpenAI’s consumer-facing ChatGPT strategy than to a pure developer API play.

Why This Matters More Than It Might Look

None of these four releases exists in a vacuum — each lab is explicitly positioning its new model against the others’ most recent release, which is exactly why they landed within days of each other instead of spread evenly across the year. That’s worth knowing because it changes how to read any single announcement: a model described as a “generational leap” in its own press release is being measured against a competitor’s model that’s often only a few weeks old, not against last year’s baseline. Reading one lab’s benchmark claims in isolation, without checking what the other three shipped in the same window, is how a genuinely incremental update ends up sounding bigger than it actually is.

What Actually Matters for a Small Business or Freelancer

Benchmark scores and parameter counts make for good headlines, but they rarely change which model is the right day-to-day pick for someone running a small store, a freelance practice, or a one-person agency. For that decision, our ChatGPT vs. Gemini vs. Claude comparison is still the more useful starting point than any single release announcement, since it’s built around actual task fit — writing, research, coding, customer support — rather than which model technically leads a given benchmark this month. If agentic, multi-step automation is the actual use case (having a model run a research task or a coding job mostly unsupervised, rather than answering one question at a time), OpenAI’s Agents API public beta is worth reading specifically for that angle, since that capability matters more for real workflows than which lab currently claims the top spot on a reasoning leaderboard.

The Realistic Takeaway

Whichever of these four ends up being the “best” model by any given benchmark next month will likely change again within weeks, given how tightly bunched these releases already are — that’s simply the pace this specific competitive window has settled into. For anyone using these tools for real work rather than following the industry as a hobby, the practical move is choosing based on the specific task at hand (long-context document work, agentic coding, everyday writing and research, or a consumer-facing assistant) rather than chasing whichever lab’s press release is loudest this week, since that ranking is genuinely likely to have shifted again by the time the next model in this same cycle ships.