INSIGHT / MODEL RELEASES

Six Frontier Releases in Ten Days: What November 2026 Changes for Product Teams

Claude Fable 5.2, Gemini 4 Carbon, Hydrogen and Neon, Qwen4-Max and GPT-6.2 shipped between 9 and 18 November 2026. Prices, access limits, expiring launch rates and benchmark caveats, and what a PM should do about them.

4 min readUpdated

Six frontier releases from four labs landed in ten days. For a product team, each one raised the same question: do we switch? The answer sits in four details the launch posts did not lead with: price per task rather than per token, launch prices with end dates, capability released only to vetted customers, and benchmark numbers whose provenance deserves a second look.

All dates and prices below come from the vendors' own release notes, model pages, rate cards and system cards, checked on 20 November 2026. Prices are USD per million input and output tokens at standard rates.

DIAGRAM
Frontier releases, 9 to 18 November 2026 Anthropic Google Alibaba OpenAI 9 10 12 16 17 18 Claude Fable 5.2 $8 / $40 Gemini 4 Carbon $2 / $10, then $4 / $20 Gemini 4 Hydrogen $1.50 / $7.50 Qwen4-Max hosted, 2.4T / 71B active GPT-6.2 Astra, Sol $8 / $40 and $2 / $10 Gemini 4 Neon $3 / $15 Input / output, USD per million tokens. Day of November on the axis.
Read diagram description

Timeline of frontier model releases, 9 to 18 November 2026. 9 November: Claude Fable 5.2 at $8 input and $40 output. 10 November: Gemini 4 Carbon at $2 and $10, rising to $4 and $20 on 1 December 2026. 12 November: Gemini 4 Hydrogen at $1.50 and $7.50. 16 November: Qwen4-Max, hosted only, a mixture-of-experts model with 2.4 trillion total and 71 billion active parameters. 17 November: GPT-6.2 Astra at $8 and $40 and GPT-6.2 Sol at $2 and $10. 18 November: Gemini 4 Neon at $3 and $15.

1. Price the task, not the token

GPT-6.2 Astra lists at $8 and $40 against GPT-6 Astra's $10 and $50: 20% cheaper per token on both sides. Whether a feature gets 20% cheaper depends on how many tokens the new model spends per completed task. Reasoning models think in tokens you pay for, and Google states that Gemini 4 output prices include thinking tokens. Measure tokens per completed task on your own traffic before updating a business case, and use cost per successful task as the headline number.

Long prompts move the bill too. Above 272,000 input tokens, GPT-6.2 Astra's input price doubles to $16 and its output price rises to $60. A retrieval design that stuffs the context window crosses that line sooner than its designers expect.

2. Launch prices have end dates

Gemini 4 Carbon launched at $2 and $10 "through November 30, 2026", rising to $4 and $20 on 1 December 2026. Gemini 4 Argon, released in October, follows the same schedule: $2 and $10 through 30 November, $4 and $20 from 1 December 2026. A cost model approved in launch week at the introductory rate will be wrong when the feature ships. Price every workload at the rate that applies on its ship date.

Other dated changes belong in the same calendar: Claude Sonnet 4.5 retires on 30 November; Gemini 2.5 Pro and 2.5 Flash requests route automatically to newer models from 1 December; GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down on 14 January 2027.

3. Some capability is released by programme, not by price

Anthropic made Claude Mythos 5.2 "available to vetted organizations that work in these dual-use areas". OpenAI's system card for GPT-6.2 reports that both models reach the "Critical" threshold for cybersecurity under its Preparedness Framework, and that Astra is the first model it has released externally at "High" for AI self-improvement. For product teams in security, biology or other sensitive domains, expect stronger safeguards and more refusals on the general endpoints, and ask about trusted-access programmes before committing a roadmap. Frontier safety frameworks explains how these thresholds become launch scope.

4. Read the benchmark table for provenance

Launch tables are company-reported, run on each vendor's harness and at named effort levels. Two signals this year argue for caution. OpenAI said in February 2026 that it no longer reports SWE-bench Verified, citing flawed test cases and contamination (as reported in its post), while Anthropic's November card still headlines SWE-bench Pro, which OpenAI later questioned too, as reported in its posts. OpenAI's own GPT-6.2 card warns that a security benchmark built from historical vulnerabilities "raises a risk of contamination that may inflate the reported results". Anthropic's card reports that Fable 5.2 recognised it might be under evaluation in as many as 43% of automated behavioural audit transcripts. None of this makes the models worse; it makes your own evaluation more important than theirs.

What to do this week

  1. Shortlist by access and price, at the price that applies on your ship date.
  2. Run your golden set on each shortlisted model at the effort setting you would deploy, and record tokens per completed task.
  3. Read the system card for refusals and safeguards in your domain.
  4. Route a slice of traffic to the best candidate with a written decision rule, following model migration.
  5. Update the dates file: introductory prices, retirements and re-routings, with owners.

The fuller method is in how to read a model release. For the architecture behind Qwen4-Max's headline numbers, see mixture of experts; for why faster tiers cost more, AI accelerators for PMs.

Questions about this topic

Between 9 and 18 November 2026: Anthropic's Claude Fable 5.2 and Mythos 5.2 (9 November), Google's Gemini 4 Carbon (10 November), Gemini 4 Hydrogen (12 November) and Gemini 4 Neon (18 November), Alibaba's Qwen4-Max (16 November) and OpenAI's GPT-6.2 Astra and Sol (17 November). Dates and prices are from each vendor's own pages, checked on 20 November 2026.