Product1 distinct publisher3 min readPublished
The new models ship with a 25% cost improvement that only lands on prompts with a stable prefix, while the capacity underwriting them does not reach the grid until early 2027. That second number is the one buyers should price.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
The 25% figure Anthropic quotes for typical workloads comes out of prompt caching [8], which means it lands on teams whose requests share a long, stable prefix and skips teams that reassemble context on every call. Agent runs resend the same system prompt and tool definitions turn after turn, which is the traffic shape a cache pays for, and the likeliest reason the higher 45% ceiling is attached to agent-heavy applications rather than to chat [8]. On a bill rather than a slide: a $40,000 monthly model spend falls to $30,000 at 25% off and to $22,000 at 45%, so the same budget buys about 1.33 times the typical volume, or about 1.82 times the agent volume [1]. Both assume the cache actually hits.
The access split is the part that ends up in a procurement questionnaire. Fable 5.1 is generally available and cannot be used for sensitive cybersecurity and biology tasks; Mythos 5.1 is effectively the same model with those blocks removed, offered to a limited number of trusted organizations through two programs, one for cybersecurity researchers and one for biologists [5][6]. The relaxed guardrails in the GA model let researchers find software vulnerabilities, but exploit-path work sits on the far side of the gate [7], and Anthropic calls Mythos 5.1 the strongest cyber model it has released [9]. This tier is built for an in-house researcher whose employer can pass a trust review; researchers without that access get the model with the exploit-path work removed.
The $35 billion is single-sourced: SiliconANGLE reports it from Wall Street Journal sources who did not say how much capacity it buys [10]. Working from Jensen Huang's estimate that a gigawatt will soon cost $80 billion to $100 billion, SiliconANGLE puts the purchase at 350 to 437 megawatts [11]. The destination is a Hut 8 campus under construction in Nueces County, Texas, rated at 1 gigawatt, where Nvidia holds the campus lease and, according to the Financial Times, is the unnamed investment-grade customer that took 704 megawatts in late July [13][14][15]. Add those and you get 1,054 to 1,141 megawatts on a 1,000-megawatt site [2], so either the extrapolation runs high or part of Anthropic's capacity sits elsewhere in Lambda's fleet, which can spin up Nvidia clusters of as many as 165,000 GPUs [12]. Less than a week earlier Anthropic signed a larger contract with Nscale that the report does not quantify [17]. Hut 8 expects grid connection in the first quarter of 2027 and the first data hall roughly six months after [16], about a year past this launch [3].
Sorting your own workloads comes down to two questions: whether the request carries a prefix long enough and stable enough to cache, and whether the capability you need sits in the generally available tier or behind the trust review. Cache-stable plus GA is the only quadrant where the quoted economics and the ship date are both yours. Unstable prefix plus GA means you budget at the old rate and book any saving as upside. Cache-stable plus gated means your roadmap turns on an access decision made by someone else. Unstable plus gated is a research project wearing a product name. Whichever quadrant you land in, the price on the page today is set against a lease whose site is not energized until 2027 [16], which matters if the term you sign runs past it.
Ranked by verification strength, evidence, and original report placement.
Anthropic PBC debuted Claude Fable 5.1 and Claude Mythos 5.1, described as its most capable large language models to date.
SiliconANGLE published its report on the launch and the Lambda deal on September 1, 2026.
The model launch came a day after Anthropic inked a $35 billion infrastructure deal with cloud startup Lambda Inc.
Anthropic says Fable 5.1 completed Terminal-Bench-Science 0.1, a test of research task ability, with a score of 52.6%, more than double what the previous-generation Fable 5 achieved.
Fable 5.1 and Mythos 5.1 are effectively the same model; the main difference is that Fable 5.1 cannot be used for sensitive cybersecurity and biology tasks, and it is generally available.
Mythos 5.1 does not block sensitive cyber and biology use cases and will be accessible only to a limited number of trusted organizations, through two specialized access programs for cybersecurity researchers and for biologists.
Distinct publishers with included, body-backed reporting in this cluster.
2 articles · September 1, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
invest
Anthropic Cuts Cache-Read Prices by 75%; Cache Reads Were ~60% of a Heavy Agent's Bill Before the Cut1 distinct publisher
build
Anthropic's $35B Lambda contract puts Nvidia on both sides of the table3 distinct publishers
build
Fable 5.1's 52.6% science score arrives on a benchmark that was five days old1 distinct publisher
leadership
Anthropic cuts Fable 5.1 prices by 25% and launches two-tier safeguard system with Mythos 5.12 distinct publishers
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One newsroom, printed twice
Both copies in our coverage are the same SiliconANGLE report, so the source count overstates the witness count. Anthropic supplies every capability and cost figure; the $35 billion arrives through people who would not be named, speaking to the Wall Street Journal; the only numbers with an identified institution attached are Hut 8's 704-megawatt disclosure and the Financial Times' identification of Nvidia as the tenant. No benchmark here was re-run by anyone outside the vendor.
Shipped; nobody using it yet on the record
What exists is availability. Fable 5.1 is out and Mythos 5.1 deliberately is not, reaching a short list of trusted labs through application programs. Not one customer, workload or spend figure appears. On the infrastructure side the record is a signed lease against a campus that has not been energized — the first data hall is a date, not a deployment.
Records now, electricity in 2027
The superlatives all belong to Anthropic: most capable to date, records across several benchmarks, the strongest cyber capabilities it has released. Against that, the coding gain is printed as '13.%' and left there, the megawatts behind $35 billion are a division problem rather than a disclosure, and the campus meant to carry the load connects to the grid in early 2027. The gap is less exaggeration than timing — the claims are present tense and the capacity is not.
Everyone in frame benefits from a big number
A capability launch lands one day after a $35 billion lease becomes public, and the lease figure comes from parties who chose anonymity. Nvidia sits on both sides of the same campus — holder of the lease, which SiliconANGLE reads as risk relief for Lambda and Anthropic, and separately the 704-megawatt anchor tenant the Financial Times named. Then SiliconANGLE signs off with its own AWS marketplace and membership pitch, worth knowing when weighing how warmly the deal is framed.
Direction firm, figures soft
That Anthropic shipped two access tiers and committed heavily to leased compute is safe to believe. Almost everything a reader would act on — the benchmark deltas, the 25% and 45% savings, the megawatt count — comes from interested parties through a single retelling. And two of the sharpest observations in this story, the budget arithmetic and the capacity overshoot, are sums we did ourselves rather than facts anyone confirmed.