Fable 5.1 Ships Cheaper. Astra Ships Gated.
Yesterday was a 35 billion Texas power contract. Overnight the labs shipped the models that contract is supposed to feed.
Anthropic put Claude Fable 5.1 on every cloud. OpenAI said Astra is the first model it will classify as Critical under its own Preparedness Framework, and it will not ship the cyber bits to everyone. Same 24 hours. Opposite product strategies.
The cache bill is the product
Fable 5.1 and Mythos 5.1 are the same weights with different locks. Anthropic’s post is unusually blunt about why customers were yelling: price, data retention, and safeguards that kept blocking homework.
List price did not move. Input is still $10 per million tokens. Output is still $50. Cache reads dropped 75 percent, to $0.25. Anthropic says that cuts typical billed work about 25 percent, and highly agentic sessions up to about 45 percent, because cache reads are most of the bill once a coding agent is chewing the same repo for an hour.
That is not a list-price cut. It is an admission that the last generation’s unit economics were cache, not intelligence.
| Lane | What shipped Sept 1 | Who gets it |
|---|---|---|
| Fable 5.1 | Same $10 / $50, cache reads $0.25, more precise cyber/bio blocks | GA on API, AWS, Google Cloud, Azure |
| Mythos 5.1 | Same model, looser cyber and life-sciences locks | Trusted access only (CVP / LSVP) |
| Astra | Critical cyber designation; public version stripped of the sharp tools | Testers first, then Daybreak Blue defenders |
On vendor benches, Fable 5.1 posts 55.8% on Terminal-Bench 4.0 versus Fable 5 at 42.0% and GPT-5.6 Sol at 37.3%. Mythos 5.1, with the cyber classifier out of the way, hits 60.9%. CursorBench 3.2.0 is 73.4%. Treat those as Anthropic’s own numbers until someone reruns them.
Cognition says it is moving Devin’s Opus 5 traffic onto Fable 5.1 on launch day because the cache math finally makes a Fable-class model cheaper than parking code review on Opus. Jane Street, MongoDB, Ramp, and Datadog all showed up in the launch quotes with the same shape: long unattended runs that used to drift now stay readable.
Enterprise Frontier Safeguards is the other concession. Customer-controlled cloud storage, customer-run review, zero-data-retention until it rolls this fall. Fable 5.1 can now find vulnerabilities. It still will not write the exploit. Cyber false positives drop about 60 percent per Claude Code session. Distillation got a new speed bump: new API accounts cannot edit prior thinking while keeping the transcript.
If you are still running the agent on a desk instead of Bedrock, a fast NVMe and a quiet UPS remain cheaper than waiting for EFS to land in your region.
Astra is Critical. The public copy is not.
OpenAI’s Path to Astra note closes the August maybe. Astra meets the Critical cybersecurity threshold. With tools and access, it can find unknown flaws and build exploits across hardened systems without a person steering each step. First model they have put in that bucket.
The receipts they published:
- 100% on ExploitBench (known-vuln exploit development)
- Two zero-days found and chained on an internal V8 set; they say they are disclosing to maintainers
- A browser-compromise chain that escaped the sandbox and ran host commands from an HTML file
- A local privilege-escalation chain from unprivileged user to root on a hardened OS
- Jailbreak refusals at 91.5%, up from 59% on GPT-5.6 Sol
TechCrunch and WIRED have the same briefing: public Astra “soon,” advanced cyber only for testers, then Daybreak Blue. Extra chain-of-thought monitoring will pause or stop tasks. ChatGPT and Codex users get a review prompt. API tasks just die.
OpenAI says Astra was not in the Hugging Face incident, and that a honeypot built from that mess got zero escape attempts from Astra versus 56% from unprotected GPT-5.6 Sol. Believe the designation. Treat the alignment graphs as vendor graphs.
Fable 5.1’s answer to the same week is the opposite polarity: let the GA model find bugs, keep exploit-writing on Mythos, and stop lighting up biology homework. OpenAI’s answer is a fence. Both labs spent August learning that agent sandboxes leak. They spent Tuesday shipping different doors.
A hardware security key will not get you Daybreak Blue. It will still beat a reused password the next time a session-stealer walks off with a cookie.
Dell just printed the rack receipt
While the labs argued about fences, Dell reported the quarter that buys the racks.
The 8-K exhibit is not subtle. Record revenue $47.0 billion, up 58%. AI-optimized servers $16.4 billion, up 100%. AI orders in the quarter $60.9 billion. Backlog exiting Q2: $95 billion. Full-year FY27 guide jumped $25 billion to $192 billion, with AI-server revenue now $74 billion.
Jeff Clarke’s line is the useful one: demand is broadening across neoclouds, sovereigns, and enterprise, and the customer count is past 6,500. Traditional servers and networking still grew 122%. The AI box is not eating the rest of the catalog. It is dragging it.
If you are buying one box instead of a $95 billion backlog, a used workstation and a gigabit managed switch still beat waiting for a neocloud to free a reservation.
The pattern for the week is now ugly and clear. Anthropic is discounting the tokens that sit in cache so Fable-class agents can run all night. OpenAI is putting the model that can write the exploit behind a named program. Dell is booking the next ninety-five billion of the buildings those agents will live in. None of that is a benchmark. It is a billing change, a gate, and a purchase order.
Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.
Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.