The New Flagships by the Numbers: Claude, GPT-6 Astra, Gemini 4 Argon and Muse Wonder

Six models in six weeks

Anthropic launched Claude Fable 5.1 on September 1, 2026, at $10 per million input tokens and $50 per million output tokens, the same list price as Claude Fable 5, with cache reads cut to $0.25.¹ OpenAI followed on September 3 with GPT-6 Astra, which it calls "our most capable model, built for the hardest end-to-end work", at the same $10/$50.⁴ Anthropic added Claude Opus 5.5 on September 22 at $4/$20, and OpenAI added GPT-6.1 Sol on September 29 at $2/$10, described as "near-Astra performance" at lower cost.¹⁴⁵⁶

Two more arrived in October. Meta introduced Muse Wonder and Muse Wonder Full on October 12, calling them its "most capable models to date"; Muse Wonder is on the Meta Model API at $4/$20, while Muse Wonder Full is available only to Muse Code subscribers, and Meta's pricing page lists no per-token price for it.¹⁰¹¹¹² On October 13, Google listed Gemini 4 Argon as generally available, calling it its "flagship reasoning model", at $2/$10 through November 30.⁷⁸⁹

Three price tiers, one ratio

The six list prices form three groups. Claude Fable 5.1 and GPT-6 Astra both list $10/$50.³⁵ Claude Opus 5.5 and Muse Wonder both list $4/$20.³¹² GPT-6.1 Sol lists $2/$10, and so does Gemini 4 Argon, but only through November 30, 2026: Google's pricing page lists $4/$20 from December 1, which would put Argon in the same tier as Opus 5.5 and Muse Wonder.⁶⁷

In all six rows the output price is five times the input price (a ratio computed from the listed prices, not stated by the labs). Cache-read prices vary more: from $0.10 per million tokens (GPT-6.1 Sol, and Gemini 4 Argon until December) to $1.00 (GPT-6 Astra).⁵⁶⁷ Anthropic, Google and OpenAI (for Batch and Flex) each list 50% off for batch-style processing, while Meta's pricing page lists no batch discount.³⁵⁷¹² Prices are per token as each lab's tokenizer counts it, so the same text can cost a different number of tokens at different labs, and a lower list price is not automatically a lower cost per task.

Context windows: all about 1M, counted differently

Every row has a context window of roughly one million tokens: 1M for the two Claude models,² 1,050,000 for the two OpenAI models,⁵⁶ and 1,048,576 for Gemini 4 Argon and Muse Wonder.⁸¹¹ The difference is in the rules around it. Google states Argon's window as a unified limit on input and output together, with a default maximum output of 65,536 tokens and up to 1,048,576 with its "long decode" setting.⁸ Anthropic and OpenAI both list a 128,000-token maximum output.²⁵⁶ Meta's documentation gives the window but not a maximum output length.¹¹

Long prompts are priced differently. Anthropic says the full 1M window is billed at the standard rate: a 900,000-token request costs the same per token as a 9,000-token one.³ OpenAI charges 2× the input and cache rates and 1.5× the output rate, for the whole request, once a prompt exceeds 272,000 input tokens.⁵⁶ Meta states there is no long-context premium,¹² and the Argon pricing section we read lists a single price.⁷

Reasoning controls and availability

Each lab exposes reasoning differently. Anthropic lists adaptive thinking as always on for Fable 5.1 and Opus 5.5.² OpenAI's Astra supports reasoning effort levels low, medium, high, xhigh and max, and not "none".⁴⁵ Google lists Argon's thinking levels as low, medium and high, and its output price includes thinking tokens.⁷⁸ Meta's Muse Wonder supports effort levels from "minimal" to "max", and setting "max" on Muse Wonder does not select Muse Wonder Full.¹¹

Availability also differs. Meta's documentation labels Muse Wonder as in preview.¹¹ Argon is not available on Google's free tier or on its AI Pro plan.⁷ OpenAI added an "ultrafast" service tier for Astra on September 29, 2026, which has its own pricing and isn't in the table.⁴

What the labs say about performance

This post ranks nothing by capability. Each lab publishes its own benchmark results, run with its own harness, and vendor-run results aren't compared with independent ones on this site. One claim is worth attributing: Meta's developer site says Muse Wonder Full "tops agentic and coding benchmarks, including DeepSWE and SWE-Marathon in independent third-party runs", a claim we haven't checked against those runs.¹³ For how coding scores are produced and why they get retired, see SWE-bench Verified, Pro and What Replaced Them, and for where these prices sit against the rest of the market, Cheapest LLM APIs per 1M Output Tokens.

Reading the numbers

Three things limit these comparisons. First, every figure is vendor-stated and was read from the lab's page on October 14, 2026; prices, promotions and limits change, and Argon's price is scheduled to double on December 1. Second, a flagship's list price says nothing about how many tokens a task consumes, which depends on the tokenizer and on how much the model reasons before answering. Third, the table covers only the first-party API: the labs also sell these models through cloud marketplaces and subscriptions at other prices, which are not compared here.