Anthropic files its public S-1, GSA gets Claude at 40% off, and the agent runtimes ship in lockstep

Anthropic, PBC says it filed a Form S-1 with the SEC on October 26 for a proposed IPO of Class A common stock on the Nasdaq Global Select Market under the ticker “ANT”, with Goldman Sachs and Morgan Stanley named as sources for the preliminary prospectus. Share count and price range are not yet set, and the registration statement is not yet effective.

Why it matters now: this is the step after the confidential draft we covered earlier — the document becomes public, which means the real numbers, compute commitments and customer concentration land in front of everyone. The announcement itself carries no financials, so treat press figures circulating before the prospectus is read as secondary until verified against the filing.

A new GSA agreement gives federal, state, local, tribal and territorial buyers 40% off Claude Enterprise, Claude for Government and the Claude API, through GSA’s Multiple Award Schedule, with no minimum spend. Anthropic estimates roughly $1 billion in government savings and says about 500 US public-sector organisations already use Claude, with federal civilian usage up about 20x in a year. Claude for Government has been FedRAMP High and generally available since September 30; Claude Enterprise is in process for FedRAMP Moderate.

Why it matters now: government procurement is where agent adoption gets locked in by contract vehicle rather than by developer taste. The figures are vendor-reported, but the Utah pilot detail — 34 developers on Claude Code, 80% productive within 30 minutes — shows the buy is being justified by agentic coding, not chat.

The expansion names 20 partners, including the Sloan, Simons and Schmidt foundations, the US National Science Foundation, the Wellcome Sanger Institute, the Francis Crick Institute, EMBL and Gladstone, plus projects such as FlyWire, gnomAD and an open Lean library. HHMI Janelia is working with Anthropic on a Model Hardware Standard, currently a research preview.

Why it matters now: this widens the August seat-and-credit programme from individual labs to institutions and funders who can underwrite agent use at scale. No credit amounts are stated, so the size of the commitment is unknown.

2.1.312 adds oidc.user_id_claim to the Claude apps gateway so enterprises can key identity on a claim other than sub (Microsoft Entra’s oid, for example), uses a shorter system prompt on Bedrock, Vertex, Foundry and Mantle for unrecognised model names, caps prompt and agent hooks at 24 running per session, and fixes Claude overwriting a very large CLAUDE.md after reading only its start. 2.1.311 adds --hold-idle-file and --max-hold-min to the self-hosted runner, fixes standard-context sessions dying on “Usage credits required for 1M context” (it now compacts and continues), and fixes SubagentStop hooks firing about twice a minute.

Why it matters now: none of this is a headline feature, and that is the point — the work is identity, runner lifecycle and hook hygiene for fleets of agents inside companies. The CLAUDE.md overwrite fix is worth reading twice if you keep long instruction files.

The stable release adds a built-in Google Cloud Vertex AI provider (model_provider = "gcp"), lists GPT-6.1 Terra in the model catalogs, makes Memory v2 the default for fresh profiles with import of external-agent memories, lets reasoning effort change mid-conversation, and closes a marketplace URL validation gap that could let Git connect to a different host. 0.166.0 alphas followed the same day.

Why it matters now: an OpenAI coding agent shipping a first-party Google Cloud provider is the clearest sign the harness is being treated as neutral infrastructure, while default memory plus memory import is a direct play for users coming from other agents.

The 0.66.0 release adds a banner saying the CLI is not yet optimised for Gemini 4, opt-in private network endpoints for MCP OAuth flows, a keys array in keybindings.json and terminal rendering fixes. The 0.67.0 preview adds a models command and passes MCP server and tool metadata into ACP permission requests.

Why it matters now: the banner is the story — it is a product-side acknowledgement that a Gemini 4 model exists to be optimised for, from the people shipping the harness. Beyond that, the MCP OAuth and ACP changes show enterprise auth moving into the tool-calling layer.

2026.10.8 restarts the Gateway when a post-update Doctor step fails instead of leaving it stopped, requires approval for test -v operand expansion in allowlist mode, refuses browser-proxy commands on node invoke, and stops two TTS providers’ errors from exposing API keys. 2026.9.35 is a gateway-only extended-stable build with heap-cap, idle Codex client and Doctor repair fixes; 2026.10.10-beta.4 is already out.

Why it matters now: the allowlist bypass and key-leak fixes are the type of bug that matters once an agent holds real credentials, and an explicit extended-stable lane is how a project signals it expects long-lived production gateways.

The default DeltaChannel snapshot interval drops from every 1,000 updates to every 100, error_handler recovery is fixed on every path, and the last checkpoint now streams correctly when a run stops at interrupt_after or the recursion limit. The SDK (0.4.10) retries runs.wait and adds durability and stream_resumable to its run overloads.

Why it matters now: a tenfold change to snapshot cadence trades storage for faster resume of long-running agents, and the interrupt fixes are what human-in-the-loop workflows depend on.