Release Date and GA Status: September 1 vs September 3-4
Claude Fable 5.1 and Claude Mythos 5.1 launched with Anthropic's official announcement dated September 1, 2026. Fable 5.1 went straight to general availability (GA); Mythos 5.1, on the other hand, is accessible only through Anthropic's trusted-access programs — meaning not every user can reach Mythos 5.1. Two days later, on September 3, OpenAI announced GPT-6 Astra and began a staged rollout: first to a limited group of organizations, then within days to ChatGPT Plus, Pro, Business, and Enterprise users, while also rolling out to the OpenAI API, Microsoft Azure, and AWS Bedrock. Two more notable developments followed on September 4: Astra became generally available on GitHub Copilot (Fable 5.1 had reached GA on the same platform three days earlier, on September 1), and Codex CLI version 0.153.4 made Astra the bundled default model whenever no explicit model configuration was set (a few hours earlier that same day, version 0.153.3 had added Astra to the Amazon Bedrock model picker). The upshot: both models are barely a week old as of September 7 — there's no long-running production usage data yet, but the pace of rollout has been strikingly aggressive for both.
Pricing and Cache Economics: Nominal Parity, the Real Difference Is in Cache
On the surface both models sit at the same price point: $10/MTok input, $50/MTok output — for standard, short-context usage. But the real story is hidden in cache-read pricing. Where Anthropic's standard models price cache-hit at 0.1x of base input, Fable 5.1 and Mythos 5.1 got a special reduction to 0.025x — that is, $0.25/MTok per cache-hit. Anthropic states this delivers ~25% savings versus Fable 5 in typical workloads, and up to ~45% in heavy agentic scenarios. On the GPT-6 Astra side, cache-hit sits at the standard 0.1x multiplier, $1.00/MTok — 4x Fable 5.1's rate; on top of that, Astra has a threshold above 272,000 tokens where input/cache pricing rises 2x and output 1.5x. Looking at unit prices alone, Fable 5.1 looks cheap in cache-heavy scenarios. But the total bill isn't just about unit price: in Artificial Analysis's per-task measurement, Fable 5.1 (max) came out to $6.12 and GPT-6 Astra (max) to $2.57 — the same task set cost roughly 2.4x more on the Fable side. The reason is tokens spent: Anthropic's own pricing documentation states that Claude 4.7 and later models generate roughly 30% more tokens for the same text, on top of which reasoning tokens are layered in. A cheap unit price doesn't mean a cheap bill.
Context Window and Knowledge Cutoff
Both models have practically equivalent context windows: Fable 5.1 at a full 1,000,000 tokens, GPT-6 Astra at 1,050,000 tokens — the ~5% difference doesn't create a noticeable advantage in day-to-day use. Output limits are also equal: 128,000 tokens for both. The real divergence is in knowledge cutoff: Fable 5.1's reliable knowledge cutoff extends to June 2026, while GPT-6 Astra's cutoff is April 30, 2026. That roughly two-month gap gives Fable 5.1 a small but real freshness advantage on things like fast-moving library versions, new framework APIs, or recent security vulnerabilities. Another point worth noting for long-context usage: Astra's price increase kicks in above the 272K-token threshold, while Fable 5.1 has no such tiered pricing structure — a single price applies up to 1M tokens.
Which Tools Default to Which: Claude Code vs Codex CLI
According to Anthropic's own launch announcement, Fable 5.1 runs at High effort by default in Claude Code; in Claude Cowork and Claude.ai the default is Medium effort. This means the model runs with a deeper thinking budget in CLI/agentic development scenarios, while it's kept at a medium level for faster responses in consumer-facing interfaces. On the GPT-6 Astra side there's a clear turning point: Codex CLI version 0.153.4 (September 4, 2026) made Astra the bundled default model whenever no explicit model configuration was set — meaning developers who hadn't specified a model in config.toml were automatically switched to Astra. OpenAI's own model catalog also puts it forward as the flagship model, saying 'use GPT-6 Astra if you don't know where to start,' and positions it on the model card as 'the most capable model built for the hardest end-to-end tasks.' The practical upshot: in both ecosystems, the flagship tool (Claude Code / Codex CLI) pushes forward its own newest model — the tool you're already using largely determines the model you end up with.
Adaptive Thinking and Effort Control
Both models have a 5-level effort/reasoning control system: low, medium, high, xhigh, and max. In Fable 5.1, thinking is described as 'adaptive (always on)' — meaning the thinking mechanism can't be fully turned off, only its intensity can be adjusted; this differs from the plain 'adaptive' label on Opus 5 and Sonnet 5, and is a stricter behavior. GPT-6 Astra's reasoning.effort parameter supports the same 5 levels but with one notable exception: the 'none' effort level isn't supported, meaning Astra too is forced to do some amount of reasoning even at its minimum level. Where Astra differentiates itself is an input item called configuration_update, which lets it raise the effort level mid-conversation without rewriting the existing prompt prefix — and therefore without breaking the cache. This offers a way to increase depth while preserving the cache advantage when a long agentic session suddenly shifts to a harder subtask.
Independent Benchmark Status: Artificial Analysis Measurements
On September 3, 2026, Artificial Analysis published the first independent analysis evaluating both models together; the comparison page updates live. The result isn't one-directional: on the Intelligence Index (v4.2), Fable 5.1 (max effort, default fallback) leads slightly at 57 points versus Astra's (max) 55; on the Coding Agent Index, Fable also leads, 70-67 (Fable measured on Claude Code, Astra on the Codex harness — the harness difference could affect the result). On the speed side the picture is nearly neck and neck: both sit around the ~70 token/s band (Artificial Analysis's live measurement, September 7, 2026 reading: Astra 67, Fable 5.1 69); but on time-to-first-answer-token (including thinking), Astra is markedly slower (370s vs 265s). In short, saying 'X won' would be misleading at this stage — the result changes depending on which metric you look at. Read the learning-curve, job-market, and future-proofing scores on this page with that lens: since both models are only a week old, there's no independent data on these axes yet; those scores are editorial.
Vendor Tables and Their Limits
Anthropic's launch table was published before Astra, so it can't include it; OpenAI's table, on the other hand, lists Fable 5.1 directly — but the benchmark selection and run conditions belong to the vendor itself, with no independent replication. On Terminal-Bench 4.0, Astra leads: 57.9% versus Fable 5.1's 55.8% in OpenAI's own table — and this second number matches Anthropic's own launch table exactly; both vendors cite the same figure. In the same table, Astra also leads on DeepSWE v1.1 (74.1% vs 67.4%), FrontierCode 1.1 Main (53.3% vs 50.9%), and OpenAI's own database-migration tasks (63.9% vs 57.8%); the gap is smaller on FrontierCode 1.1 Extended (64.5% vs 63.6%). The table's footnote states the scores are the 'maximum across any effort' level and that the GPT measurements were run in OpenAI's own research environment — meaning the figures were taken under conditions the vendor chose. In this table, Fable 5.1's Coding Agent Index cell is empty; Fable's score of 70 comes from Artificial Analysis's own article.