Three flagship model launches landed within four days of each other in early September: Claude Fable 5.1 and Mythos 5.1 from Anthropic on 1 September, Gemini 3.8 Flash and 3.8 Flash Cyber from Google on 2 September, and GPT-6 Astra from OpenAI on 3–4 September. Taken individually, each is a routine capability update. Taken together, they show the same structural decision made three times independently: ship a general model to everyone, and gate the sharpest edge of its cybersecurity capability behind a restricted programme.
What each lab actually shipped
Claude Fable 5.1 is generally available on all platforms and keeps Fable 5's headline pricing of $10 per million input tokens and $50 per million output tokens. The change is in cache-read pricing, which drops 75% to $0.25 per million tokens from $1. Anthropic estimates that saves around 25% on typical workloads and up to 45% on complex, highly agentic tasks — the ones that reuse a long system prompt or context window repeatedly. Mythos 5.1 is the same underlying model with fewer safeguards, and is available only through trusted access programmes aimed specifically at cybersecurity and life-sciences work.
Gemini 3.8 Flash is Google's third Flash-tier release in six weeks, priced at $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026, rising to $1.50 and $7.50 from January 2027. It scores 90.8% on Terminal-Bench 2.1, up from 81.6% for the previous Flash model. Gemini 3.8 Flash Cyber, a companion variant tuned for vulnerability detection and patching, is restricted to Google's Fairwind Program for governments and trusted partners and is reported to find real-world vulnerabilities at a rate above 70%.
GPT-6 Astra rolled out first to a limited set of organisations in OpenAI's application-based cybersecurity programme, then to ChatGPT Plus, Pro, Business and Enterprise users within days, plus the API and AWS. It is priced at $10 per million input tokens and $50 per million output tokens — 2.5 times the current promotional rate for GPT-5.6 Sol. OpenAI states Astra is the first model to reach the Critical threshold for cybersecurity capability under its Preparedness Framework, meaning it can, with the right tools, find and exploit previously unknown security flaws largely unsupervised. The public release ships with restrictions on cybersecurity-related prompts by default; the unrestricted capability stays behind the same kind of vetted access programme as Anthropic's and Google's variants.
Why the pattern matters more than any one launch
Three competing labs independently reached the same structure in the same week: a general-purpose model available to anyone with a subscription or API key, and a second, more capable tier reserved for vetted defenders. That is a signal about where the industry now believes the genuine risk sits — not in the general reasoning and coding capability, which keeps shipping broadly and getting cheaper, but specifically in autonomous vulnerability discovery and exploitation.
For a small team, the practical takeaway is that you will very likely never get direct access to the sharpest tier of any of these models. Do not plan a product around eventually reaching it. What you can plan around is the general-release models, which are genuinely more capable and, in Google's and Anthropic's cases, cheaper for how most SaaS teams actually use them.
Which one to actually reach for
Gemini 3.8 Flash is the one to default to for high-volume, low-complexity work — classification, extraction, simple agent steps — where its low per-token cost and fast iteration cadence from Google matter more than frontier reasoning.
Claude Fable 5.1 is worth the switch if your workload is agentic or coding-heavy and reuses context across calls; the 75% cache discount compounds fast on anything that keeps a large system prompt resident.
GPT-6 Astra is priced at a real premium over the tier it replaces, and unless your task specifically needs frontier-level reasoning or you are one of the teams doing legitimate defensive security research applying for the vetted cybersecurity programme, it is worth testing against Fable 5.1 on your actual workload before committing budget to it.
The bottom line
Three labs, one week, the same structural answer: broad access to a stronger general model, narrow access to its cybersecurity edge. Route routine, high-volume work to Gemini 3.8 Flash, route context-heavy agentic and coding work to Claude Fable 5.1, and treat GPT-6 Astra's 2.5x pricing as a cost you pay only when the task genuinely needs it.