May opened with a developer conference and closed with a $965 billion valuation. In between, Anthropic signed 300 megawatts from SpaceX, signed $1.8 billion with Akamai, acquired the company that generates SDKs for OpenAI and Google and then announced it would shut that service down, shipped a flagship model at the previous flagship's price, and disclosed that roughly fifty partners had found more than ten thousand high-or-critical vulnerabilities using a model the public could not have.
Read as a list of releases, May looks ordinary: one model, about twenty Claude Code patches, a steady SDK cadence. Read as a set of purchases, it is the month Anthropic stopped behaving like a company that ships models and started behaving like one assembling a supply chain. Compute, tooling, distribution and capital, all four bought in twenty-five days.
This is the archive issue for May 2026, compiled in August from the daily sweeps. It is the first Shipped. Monthly published for a month that had already closed.
The compute came first. On May 6, alongside its first developer conference, Anthropic announced that it had signed SpaceX for all AI capacity at Colossus 1 in Memphis: more than 220,000 NVIDIA GPUs drawing over 300 megawatts. Two days later Bloomberg reported a $1.8 billion, seven-year infrastructure contract with Akamai, the largest deal in Akamai's history. Fortune put the demand side next to it: 80-fold growth in annualized revenue in Q1 2026.
Then the toolchain. On May 18 Anthropic acquired Stainless, the startup that generates production-ready SDKs from API specifications, and said it would wind down all hosted Stainless products including the generator itself. Stainless built every official Anthropic SDK. It also built the Python, TypeScript, Go, Java and Kotlin SDKs for OpenAI, Google, Cloudflare and Meta. Existing customers keep what they have already generated, with full rights to modify. The factory closes; the parts stay where they landed.
That is the move worth sitting with. Buying the SDK generator does not make Anthropic's own SDKs better in any way it could not have bought as a service. What it does is remove a shared dependency from underneath every competitor at once, at a price The Information reported in the low hundreds of millions. Infrastructure you can rent is a cost. Infrastructure your rivals rent from the same vendor is a position.
Then the capital. On May 29 Anthropic closed a $65 billion Series H led by Altimeter, Dragoneer, Greenoaks and Sequoia at a $965 billion post-money valuation, surpassing OpenAI as the most valuable private AI company, on run-rate revenue of $47 billion. The trajectory is the part to keep: $9 billion ARR at the end of 2025, $14 billion at the Series G in February, $30 billion in April, $47 billion in May. The round had been reported since April 29 at a targeted $50 billion and an $880 to $900 billion valuation. Both numbers came in above the reporting.
Claude Opus 4.8 shipped May 28 at Opus 4.7 pricing: $5 and $25 per million tokens, 1M context on Claude API, Bedrock and Vertex, 200k on Foundry, 128k max output. SWE-bench Pro went to 69.2% from 64.3%, SWE-bench Verified to 88.6% from 87.6%. USAMO 2026 jumped to 96.7% from 69.6%, which is a 27-point move in one release.
The number Anthropic led with was none of those. Opus 4.8 scored 0% on uncritically accepting and reporting flawed results, the first Claude model to reach that mark, with a more than tenfold reduction in overconfidence against 4.7. That is a benchmark about whether the model will tell you your premise is wrong. For anyone running agents unattended it is worth more than the coding delta, because a coding delta makes a good agent faster and an honesty delta makes a bad run stop.
Fast mode also dropped to 2x standard pricing for 2.5x speed, down from 6x. Worth noticing for anyone whose bill moved in May without their usage moving.
Anthropic published its first formal update on Project Glasswing on May 25: roughly fifty partners found more than 10,000 high-or-critical-severity vulnerabilities using Claude Mythos Preview in the program's first month. Cloudflare reported 2,000 bugs, 400 of them high or critical, at a false-positive rate better than human testers.
Anthropic's own summary of the problem is the honest part: the relative ease of finding vulnerabilities compared with the difficulty of fixing them amounts to a major challenge for cybersecurity. Ten thousand findings in a month is a defender-first number only if the patch rate can follow it. Nothing published in May said it could, and that gap is the whole argument for who gets the unrestricted model.
On May 25 a claude-mythos-1-preview identifier also briefly appeared in Claude Code before Anthropic pulled it.
A 1:1 mirror of the month's confirmed Anthropic releases, May 06 to May 31, 2026. Use it as reference. Share it with your team.
Daily digest at 9 PM ET. Weekly magazine every Friday morning. Six labs, one feed. No spam, one-click unsubscribe.