Anthropic ships Fable 5.1 and Mythos 5.1 — unambiguously state-of-the-art and pitched as much on price, data retention, and safeguards as raw capability. But real-world costs are messier than the blog post, and the real question isn't whether to switch anymore: it's where the new model fits in your stack. Plus, OpenAI says Astra has crossed its critical cybersecurity threshold.
Stop asking whether to switch models. Ask where the new one fits.
Fable 5.1 is the new state of the art on essentially every benchmark, and Anthropic is selling it as much on cost cuts, zero data retention, and better safeguards as on capability. But with token-hungry defaults blowing through usage limits, the old question — is this good enough to switch to? — is obsolete. The best users are building a personal model architecture: matching each model to the tasks where it's genuinely better, testing against their own standing benchmarks, and refusing the delusion that generally capable models are interchangeable.
Terminal Bench 4.0: 55.8 (60.9% for Mythos 5.1), up from 42% for Fable 5 and far above GPT-5.6 Sol's 37.3%. CursorBench 3.2.0: 73.4% vs Sol's 67.2%. On GDPVal AA it beat Fable 5 by 130 Elo points, leaving GPT-5.6 Sol over 140 points behind, and Automation Bench nearly doubled to 31.4%. Anthropic now holds the top three spots on the Artificial Analysis index.
Anthropic's own charts show Fable 5.1 scoring higher AND cheaper than Fable 5 at every effort level, driven by reduced cache-read pricing: an estimated 25% less for typical token-billed workloads and up to roughly 45% for highly agentic work. Even a purist lab isn't immune to the new reality that releases are judged on efficiency, not just capability jumps.
AA measured $3.76 per task versus $3.14 for Fable 5, blaming 70% higher token consumption that swamped the $1.40-per-task cache savings. The useful workaround: running on extra-high instead of max cut costs 28% for only a one-point drop in overall score.
Fable 5.1 scored 90% on ARC-AGI-2 and 97.5% on ARC-AGI-1, with average per-task cost about 32% lower than Fable 5 thanks to better token efficiency. ARC-AGI-3 results couldn't be completed: Anthropic's systems kept misclassifying the requests as reverse-engineering attempts.
Fable 5.1 scored 52.6% on Terminal Bench Science — nearly double the previous high of 29% from Opus 5. Combined with the prominent placement of agentic scientific research in the announcement, it's more evidence Anthropic wants to plant its flag in medicine, biology, and scientific research.
The new Enterprise Frontier Safeguard system lets Anthropic offer zero data retention — directly addressing the 30-day retention policies that blocked many enterprises from using Fable 5 after it came back online from its government shutdown. EFS rolls out in phases starting later this fall, but eligible customers can use Fable 5.1 with zero data retention in the meantime.
Anthropic claims an 85% reduction in fallbacks on biology and basic medicine questions and 60% fewer false positives in cybersecurity. The cyber refocus is conceptual: Fable 5.1 can be used to discover vulnerabilities without being able to develop exploits for them — and users like Matthew Miller report it patching vulnerabilities Fable 5 refused to touch.
Claude Code creator Boris Cherny says the team heard the feedback on AI-writer tells and patronizing tone, with 'solid progress' in 5.1 and more coming. Ethan Mollick's early-access verdict: a real advance in long-run work requiring judgment and taste, but less of an advance on the Claudish.
Dan Shipper calls it the strongest coding model they've used, but now fast, token efficient, and 'actually speaks like a normal person.' On agentic tasks it used about half the tokens of Opus and delivered in about 60% of the time. The old knock — a super genius in a data center that was almost unusable — appears solved.
Shipper still uses ChatGPT more day-to-day but burns way more tokens in Fable 5.1, sending it off in the morning on big end-to-end builds and checking in occasionally. That's the emerging pattern: GPT-5.6 models in Codex for interactive co-working, Fable models for long-running tasks that don't need much interaction.
Users report burning through the 20x Claude Max plan in as little as an hour with sub-agents running, and even non-hyperbolic voices like Jeffrey Emanuel blew through five-hour limits for the first time ever just auditing projects. Some call the model unusable for extended work under current rate limits.
Adam B. Levine traced much of the token burn to Fable 5.1 deciding every sub-agent should also be a Fable 5.1, ignoring long-standing rules to the contrary. On default settings, big workflows with ten-plus 5.1 sub-agents can eat even a 20x limit — a configuration fix, not necessarily a model problem.
People always price a model before anyone has figured out how to use it or the norms have settled — your mileage will likely go farther than the day-one complaints suggest. That said, Jan Velick probably has it right: subscriptions will end, and API pricing is awaiting us. At this point that feels pretty inevitable.
The idea that because multiple models can complete a task they're all interchangeable is like saying if two people can do the same work task, it doesn't matter who does it. There are tasks where that's true — and those are exactly what you optimize with cheaper models — but for high-end important work, the differences between models remain massive.
They don't have to be anyone else's tasks — research, writing, strategic thinking, building — just the things that matter to you. Especially for writing and strategy, preference is subjective: no lab can publish a benchmark for iterating on your particular mad ideas, and models others complain about may work great for your tasks.
Like enterprises building multi-model architectures, individuals should know which model and setting to use for which request rather than burning everything on the most expensive frontier model at max settings. One qualification: if you can't afford to shift between models, the 'they're all generally capable' advice holds — it's never been a better time to be locked into one ecosystem.
In a Tuesday blog post, OpenAI said its forthcoming Astra meets the critical cybersecurity capability threshold under its preparedness framework — meaning the model can find and exploit previously unknown security flaws without human guidance. It's the same concern that had Anthropic keep Mythos under lock and key earlier this year.
Astra scored 100% on ExploitBench, so OpenAI built an internal version from 20 recently disclosed high-severity vulnerabilities that couldn't be in training data. Astra hit a 30% arbitrary-code-execution rate at 40,000 tokens, while GPT-5.6 Sol couldn't produce significant results until around 110,000 — and Astra discovered and used two zero-day vulnerabilities in an exploit chain during the eval.
Additional training pushed Astra's cybersecurity refusal rate to 91.5%, up from 59% for GPT-5.6 Sol. OpenAI is also adding cyber-abuse and jailbreak classifiers, flagging higher-risk accounts for stricter guardrails, and running chain-of-thought monitoring to catch misaligned actions early. Sources suggest the model could ship as soon as this week.
— Sam Altman, on X. In an unusually serious post, Sam Altman said Astra has been done with training for a while, that models after it are being slowed as needed for safety and alignment work, and that the tension between excitement and anxiety is 'still discordant for us.' The bet: an iterative loop where society and the technology evolve together is the highest-probability path to getting this right.
The technique uses a looped transformer to process the same text multiple times before generating output, improving performance and reducing cost. The downside: part of the reasoning happens inside the model without producing output, meaning chain of thought is partially obscured and unreadable by humans. OpenAI sources say the technique was used in a limited way to keep reasoning adequately monitorable.
The worry isn't OpenAI's limited use — it's that other labs will adopt the technique without the limits. Encode AI's Nathan Calvin warned it may be hard to avoid a race to the bottom if others trade monitorability for efficiency; Redwood Research's Ryan Greenblatt fears scaling opaque reasoning until models reason almost entirely in latent space; former OpenAI researcher Steven Adler called it a violation of one of the industry's few red lines.
— Jakub Pachocki, OpenAI chief scientist. OpenAI's chief scientist pushed back: the depth of the computation graph for current frontier models including Astra is within a factor of two of GPT-4, and OpenAI has worked to preserve chain-of-thought monitoring since its first reasoning models. He conceded the technique is fragile and 'trending in a negative direction' for reasons unrelated to architecture — a topic he says he'll write about soon.
The Wall Street Journal reports internal Google engineers preferred the forthcoming 3.8 Flash to Anthropic's Opus for coding — a capability where Google has drastically lagged. Behind the scenes: every internal Gemini 3.5 Pro candidate was scrapped for not sufficiently beating the Flash models, but researchers are pleased with Gemini 4's pre-training evals, with post-training still underway.
Atlas is billed as the first multimodal world model — generating image and video frames with pixel-perfect camera control and reconstructing them in 3D, from as few as one input image. Fei-Fei Li points to use cases from VFX to robotics, and more than a few observers found it more exciting than Fable 5.1 itself, with Elvis on X calling it the most exciting release of the year.