Anthropic spent most of 2026 competing with OpenAI and Google on raw capability. Its newest release makes the case that cost is now just as important a battleground. On September 1, the company launched Claude Fable 5.1 alongside a restricted sibling called Claude Mythos 5.1, and buried in the announcement was a pricing change that could matter more to developers than any benchmark chart: cached input tokens are now four times cheaper than they were a week earlier.
That single number, a cut from $1.00 to $0.25 per million cached input tokens, lands at a moment when every major AI lab is fighting to make its models cheap enough to run constantly in the background of other software, not just answer one-off questions.
Quick facts
- Cache read pricing cut 75%, from $1.00 to $0.25 per million tokens
- Standard input and output pricing unchanged at $10 and $50 per million tokens
- Scores 52.6% on Terminal-Bench-Science 0.1, more than double Fable 5’s 24.7%
- Scores 55.8% on Terminal-Bench 4.0, up from 42.0% for Fable 5
- Mythos 5.1 is a restricted-access twin for vetted cybersecurity and life sciences work
Two models, one underlying system
Fable 5.1 and Mythos 5.1 are not two different architectures. Anthropic describes them as the same underlying model shipped with two different safety postures. Fable 5.1 is the generally available version, wrapped in Anthropic’s standard production safeguards and open to any paying customer through the API. Mythos 5.1 strips back some of those guardrails and is only available to vetted organizations working in cybersecurity research or life sciences, the kinds of users who need a model willing to engage more directly with dual-use technical material, under contractual and access controls that ordinary API customers do not get.
It is a similar pattern to how other labs have started tiering access around risk rather than around price alone. OpenAI took a comparable approach when its Astra model crossed a critical cybersecurity safeguard threshold earlier this month, restricting the sharpest capabilities to organizations that can be vetted while keeping a safer default version broadly available.
Why the cache pricing change is the real headline
Most conversations about model pricing focus on the sticker price per million tokens for a fresh prompt. In practice, a huge share of real world AI spending goes toward re-reading context that barely changes between requests, a coding agent re-scanning the same repository, a customer support bot re-reading the same policy documents, a research agent re-loading the same set of sources on every step of a long task. Prompt caching lets a model skip reprocessing that unchanged material and charge less for reading it back.
Anthropic’s new cache read price of $0.25 per million tokens is just 2.5% of Fable’s standard input price, compared with the 10% multiplier most other Claude models use for cached reads. Anthropic says that difference translates into roughly 25% lower costs for typical workloads and up to 45% lower costs for agentic workflows that lean heavily on repeated context, exactly the kind of long running, tool using agents that have become the default way developers deploy Claude in production this year.
| Pricing (per million tokens) | Fable 5 | Fable 5.1 |
|---|---|---|
| Standard input | $10.00 | $10.00 (unchanged) |
| Output | $50.00 | $50.00 (unchanged) |
| Cache read | $1.00 | $0.25 |
| 5 minute cache write | $12.50 | $12.50 (unchanged) |
| 1 hour cache write | $20.00 | $20.00 (unchanged) |
Cache pricing, not raw token cost, is quietly becoming the main lever labs use to compete on price. Photo via Pexels.
The benchmark story: research, not just chat
Anthropic is positioning Fable 5.1 less as a better chatbot and more as a better autonomous worker. The headline number is Terminal-Bench-Science 0.1, a benchmark that tests whether a model can run something close to a real scientific investigation end to end: plan the experiment, execute it in a terminal environment, interpret the results, and check its own conclusions rather than stopping at the first plausible answer.
| Model | Terminal-Bench-Science 0.1 | Terminal-Bench 4.0 |
|---|---|---|
| Claude Fable 5.1 | 52.6% | 55.8% |
| Claude Fable 5 | 24.7% | 42.0% |
| Claude Opus 5 | 29.0% | N/A |
| GPT-5.6 Sol | 22.4% | N/A |
More than doubling a score in a single model generation is unusual even by the pace frontier labs have kept up this year. It suggests Anthropic’s recent training focus, visible already in how Claude Opus 5 performed once wrapped in specialized tooling, is increasingly aimed at long horizon, self-checking agentic work rather than short single-turn answers.
The wider price war backdrop
Anthropic’s move does not happen in a vacuum. It follows months of aggressive discounting across the industry, including OpenAI cutting its own prices by 80% only for DeepSeek to undercut it again, and DeepSeek’s pricing strategy that helped push its own valuation past $70 billion despite selling access to its models for a fraction of what larger labs charge. Every lab in this race is making the same bet: usage grows fastest when the cost of leaving an AI agent running in the background approaches zero, and cache pricing is the least visible but most effective knob for getting there.
For developers already building on Claude, the practical takeaway is simple. Nothing about switching models or rewriting prompts is required to benefit from the new pricing, existing caching setups pointed at Fable automatically get the lower rate. For everyone else watching from outside, the more interesting signal is what Anthropic chose to compete on this time. Not a flashier demo, not a bigger context window, but the unglamorous cost of the tokens a model rereads a thousand times a day without anyone noticing.
Sources and further reading
- Anthropic: Introducing Claude Fable 5.1 and Claude Mythos 5.1
- VentureBeat: Fable 5.1 cuts cache read cost by 75%
- MacRumors: Anthropic launches Claude Fable 5.1
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
