AI Agent Change-Intelligence Roundup — Week of September 25, 2026

Eleven change-impact and engineering articles shipped this week. Three themes carried it: Anthropic made Opus a default (Claude Opus 5.5 launched at $4/$20 with 1M-token context, Claude Code flipped Pro and Team Standard plans to it), Anthropic re-priced the cost of refusal (the platform now bills pre-output refusals in three specific categories on every Claude plan, every platform), and the Claude API was unavailable for eighty minutes mid-week (Opus 5, Fable 5, Fable 5.1, Mythos 5, and Mythos 5.1 elevated errors between 02:50 and 04:10 UTC on Sept 22 — the third incident in twelve days). Plus OpenAI Codex rust-v0.157.0 hardened network and MCP trust boundaries and quietly removed the ultrafast tier on GPT-5.6 Sol; two Claude Code security sweeps shipped three days apart (v2.1.281 and v2.1.282); OpenClaw shipped atomic updates then had to rebuild the macOS DMG four hours later; and the build pipeline that had been wedged at the Next.js 16 webpack race all week finally recovered on Sept 25 at 15:04 UTC.

All items below cite primary sources. Prices, model IDs, version numbers, and changelog text are surfaced verbatim from the documentation they came from. Where a claim is inferred, it is labeled.

1. Claude Opus 5.5 launches at $4/$20 with 1M-token context (Sept 22) and Claude Code defaults to it

• What changed. Anthropic launched Claude Opus 5.5 (claude-opus-5-5) on Sept 22 with a 1M-token default context window, 128k max output, and always-on adaptive thinking, priced at $4 input / $20 output per MTok — the same input price as Sonnet 5 multiplied by two, and 20% cheaper than Opus 5's $5/$25 baseline. Cache hits and refreshes are $0.20 / MTok (the cheapest cache price Anthropic has ever offered, tied with Sonnet 5). At the same time, Claude Code v2.1.280 (Sept 22) made Opus 5.5 the default model on Pro and Team Standard plans, matching Max / Team Premium / Enterprise, which already had Opus by default. Four breaking API changes shipped: thinking: {"type": "disabled"} returns 400; tool_choice: {"type": "any"} and {"type": "tool"} return 400; on Claude API and Google Cloud, requests that declare a computer_20251124 tool return 400 (migrate to computer_toolset_20260801; Bedrock is unchanged); and thinking-block binding is per-model — Opus 5.5 reads blocks from Opus 5, Sonnet, and Haiku but not from Fable or Mythos (the reverse direction is also narrow). Fast mode is a research preview on Claude API only with the fast-mode-2026-02-01 beta header. Token inflation is inherited from the 4.7+ tokenizer family (so plan on ~30% more tokens vs. Sonnet 4.6 text).
• Who is affected. Every agent stack on the Anthropic stack. Teams on Pro or Team Standard who have not pinned a model in .claude/settings.json or via --model will start billing at Opus 5.5 rates on the next session. Operators migrating an Opus 5 fleet must update their request shapes for the three 400 errors. Multi-model agents that hop between Opus and Fable or Mythos have a thinking-block binding change to evaluate.
• Cost or limit impact. Marginal $1/MTok on input and $5/MTok on output vs. Opus 5 — a 20% reduction at the list price. The cache-read multiplier (Opus 5.5's $0.20 vs. Opus 5's prior $0.50 with the same 0.05x ratio; pricing page footnote 2 confirms 0.05x the base input price) cuts the marginal cost of cache-heavy agent loops by 2.5x on the Opus tier, putting Opus 5.5 in cache-price parity with Sonnet 5. For a representative long-running agent task that reads 500k cached input tokens and writes 20k output tokens per turn, marginal cost is roughly $0.10 + $0.40 = $0.50 per turn before any non-cached input — vs. Opus 5's $0.75 per turn — a 33% saving on that workload shape (list prices only; excludes prompt-cache writes and tool-result tokens). Treat the $0.50 figure as a lower bound. (Calculations by Mr. Technology from the Anthropic list prices; not measured on a live trace.)
• What could break. The Pro / Team Standard default flip is the biggest silent-cost risk: a team that was optimising around Sonnet pricing will see cost reports rise without any model name change in their config. Three request-level errors will break any agent that pinned thinking: {"type": "disabled"}, used tool_choice: "any" or "tool", or attached computer_20251124 on Claude API or Google Cloud. Multi-model agents that move from Opus 5.5 to Fable or Mythos (or vice versa) on different turns will lose thinking-block continuity. Bedrock users are unaffected on computer_20251124 but otherwise see the new model pricing immediately.
• What to do. Pin your Claude Code default model before team members hit Enter on a new session if you are on Pro or Team Standard. Audit your request shapes for the three 400-causing field-value combinations. If you are on Claude API or Google Cloud, migrate the computer-use path from computer_20251124 to computer_toolset_20260801 before flipping any agent to Opus 5.5. Update cost models: replace Opus 5 rates ($5/$25, $0.50/MTok cache) with Opus 5.5 rates ($4/$20, $0.20/MTok) in any forecast. Validate cross-model thinking-block handoffs if your agent hops between Opus and Fable or Mythos. (Inference: the cost calculations are derived math, not measured on a live trace — label as inferred.)
• Primary sources. Anthropic platform release notes, September 22, 2026 entry — verified 2026-09-23 12:09 UTC; Anthropic pricing page — Opus 5.5 row at $4 / $5 cache writes 5m / $8 cache writes 1h / $0.20 cache hits / $20 output, footnote 2 confirming the 0.05x cache multiplier; What's new in Claude Opus 5.5; Claude Code v2.1.280 release notes.

2. Anthropic bills pre-output refusals in three categories on every Claude plan, every platform (Sept 24)

• What changed. Anthropic expanded which refusals are billed at the platform level. As of Sept 24, refusals that arrive before any output when stop_details.category is bio, frontier_llm, or reasoning_extraction are now charged like any other request, at the rates of the model that ran it. Mid-stream refusals were already billed. Refusals before any output in other categories still are not billed, and fallback credit is unchanged. The change applies on all platforms (Claude API, Bedrock, Google Cloud, Microsoft Foundry).
• Who is affected. Heavy eval / red-team operators in biotech research, AI safety eval, frontier-model evaluation, and prompt-extraction red-teaming. Operators running agent workflows that intentionally probe bio, frontier_llm, or reasoning_extraction categories from Fable 5, Fable 5.1, Opus 5, or Opus 5.5 (the four models documented on Anthropic's Refusals and fallback page as including the safety classifiers).
• Cost or limit impact. Anthropic describes these three stop_details.category values as "low volumes of false positives" in the release-note entry, so most general traffic should not see any change. For evals that intentionally probe these categories — exactly the workloads that exist to generate pre-output refusals — the marginal cost of generating a refusal in one of these three categories is now equivalent to a normal request. Inferred: for a high-volume eval that explicitly maps the refusal boundaries in one of these three categories, monthly spend can rise by an amount proportional to the volume of refused requests. No measured number is available from primary sources; treat any specific percentage as inferred.
• What could break. Eval pipelines and red-team harnesses whose budget models assumed pre-output refusals in bio, frontier_llm, or reasoning_extraction were unbilled. Reporting dashboards that currently separate "billable vs. free" refusals. Auto-mode agents in Claude Code that route through these categories and rely on the off-the-bill assumption for refusal reconnaissance — Claude Code v2.1.278 (Sept 18) already moved auto-mode to server-side classifier with no separate auto-mode charge; the platform-level refusal-billing expansion is independent of that change but compounds with it for any team running auto-mode eval work.
• What to do. Audit your eval pipeline's pre-output refusal rate by stop_details.category. If you probe bio, frontier_llm, or reasoning_extraction, expect the billing line to grow and re-baseline the eval budget. Treat refusal counts in these three categories as billable going forward. If you run an auto-mode eval pipeline on Claude Code, confirm whether the cost is on the auto-mode side (handled by v2.1.278) or the platform-refusal side (this change); the two are independent surfaces. Inferred: check platform-specific pricing pages on Bedrock / Vertex / Foundry for the same refusal-billing surface; not re-fetched for this roundup.
• Primary sources. Anthropic platform release notes, September 24, 2026 entry — verified 2026-09-24 20:08 UTC; Anthropic Refusals and fallback page — documents the stop_reason=refusal / content=[] response shape, stop_details.category taxonomy, and the four affected models (Fable 5.1, Fable 5, Opus 5.5, Opus 5).

3. Claude API outage Sep 22 — Opus 5 / Fable 5-5.1 / Mythos 5-5.1 elevated errors for 80 minutes

• What changed. On Sept 22, between approximately 00:50 UTC and 02:10 UTC (Anthropic's published windows: 5:50 PM PT → 7:10 PM PT, resolved), the Claude API returned elevated error rates on Opus 5, Fable 5, Fable 5.1, Mythos 5, and Mythos 5.1. Anthropic Status listed the incident as resolved during the Sept 22 window. Two independent financial-news outlets (Hindustan Times and Moneycontrol) reported the outage the same day, characterising it as a meaningful disruption to agent traffic routed through Anthropic's top tier. Inferred: calling the window 80 minutes is a derived number (02:10 - 00:50 = 1h20m), not a separately published figure. Treat the length as inferred.
• Who is affected. Agent stacks routing production traffic through Claude Opus 5, Fable 5 / 5.1, or Mythos 5 / 5.1 on Sept 22 between roughly 00:50 and 02:10 UTC. Anyone with no fallback tier — or a fallback that lands on a lower-cost tier where latency or quality assumptions break — saw retries, timeouts, or queue back-pressure.
• Cost or limit impact. Anthropic did not announce a billing-side credit for the incident in the change-intelligence evidence chain referenced for this roundup. Inferred: most agents will absorb the cost as operational queue backup (retries, timeouts, time spent waiting on the fallback), not as a direct line item on the Anthropic invoice. Verify against the Anthropic Status post-incident summary if one was published; not independent-verified for this roundup.
• What could break. Single-tier agent fleets that route everything through Opus 5 saw elevated error rates for the duration. Multi-tier fleets with a configured fallback to Sonnet or Haiku saw traffic shift, which can surprise dashboards that pin a specific model per trace. Status-page webhook consumers that did not handle the rapid resolve / re-open cycle saw alert noise. Teams that do not run a Status-page webhook listener discovered the outage by failure rather than by being told.
• What to do. Wire the Anthropic Status page webhook into your agent control plane if you have not already. Audit the tier-by-tier fallback chain on every agent you operate and test that the fallback actually exercises on the fallback rate-limit / error path. Track the incident cadence — this is the third Anthropic status-page incident in twelve days; "rare event" assumptions are stale. Inferred: the third-in-twelve-days framing comes from the Moneycontrol republish, not from Anthropic's own page; treat the cadence as reported, not as independently tabulated.
• Primary sources. status.claude.com — Sept 22 resolved entry, 00:50–02:10 UTC; Hindustan Times and Moneycontrol republishes of the same outage (cross-referenced in the change-intelligence coverage at https://mr.technology/payloads/claude-api-outage-opus-fable-mythos-elevated-errors-sep-22-2026).

4. OpenAI Codex rust-v0.157.0 hardens network and MCP trust boundaries (and quietly removes the ultrafast tier on GPT-5.6 Sol)

• What changed. OpenAI shipped Codex rust-v0.157.0 on Sept 25 with four security-relevant changes. PR #47130 removed service_tier: "ultrafast" from gpt-5.6-sol. PR #47389 and PR #47407 closed a network restriction enforcement gap that previously allowed redirects and WebSocket upgrades to slip past the configured allowlist. PR #47094 restricted Unix-spawned local MCP servers to the stdio transport only (no other transport when launched as a local subprocess). PR #47132 added caller-provided certificate authorities to the network proxy so MITM interception can be set up explicitly. The release also fixed the network proxy certificate store to honor caller-provided certificates.
• Who is affected. Every Codex operator, with the heaviest impact on: (a) production Codex flows that depended on service_tier: "ultrafast" on gpt-5.6-sol for latency-sensitive paths (the tier is gone — those flows must move to a different tier or a different model); (b) anyone enforcing a network allowlist against Codex (the gap was a real bypass; flows that depended on it must be reconfigured); (c) MCP-using Codex flows whose Unix local MCP server previously could be reached over a non-stdio transport (now stdio only — confirmed with primary-source quotes around PR #47094).
• Cost or limit impact. The removal of the ultrafast tier on gpt-5.6-sol is a compatibility / limits change, not a price-per-token change: callers that explicitly requested ultrafast on that model now get a different tier or an error. Inferred: pricing for the standard tier on gpt-5.6-sol is unchanged; treat that as inferred (not independently re-fetched for this roundup). The network-proxy MITM-CA support is capability / debugging, not pricing.
• What could break. Production flows pinned to service_tier: "ultrafast" on gpt-5.6-sol break on the next request (the call returns an error; the exact error code is not independently verified for this roundup). Code that relied on redirects or WebSocket upgrades passing a network allowlist silently now blocks. Unix local MCP servers that previously exposed a non-stdio transport (e.g. a unix-socket or HTTP-spawned variant) can no longer be reached over that transport — only stdio.
• What to do. Audit your Codex configs for explicit service_tier: "ultrafast" and migrate those flows before upgrading Codex to rust-v0.157.0. Re-evaluate network allowlists on Codex: if a connection path depended on a redirect or WebSocket upgrade crossing an allowlist, fix the allowlist or the path. If you run a local Unix MCP server and you depend on a non-stdio transport, plan a migration to a stdio-only architecture. Set up a caller-provided CA explicitly if you actually want MITM interception (per OpenClaw precedent — leaving this unset is the safe default). Inferred: "Exact error code when ultrafast-pinned gpt-5.6-sol call is made" — not independently verified; treat as inferred.
• Primary sources. GitHub release tag rust-v0.157.0; PR #47130 (ultrafast removal); PRs #47389, #47407 (network restriction enforcement); PR #47094 (MCP stdio restriction); PR #47132 (MITM CA support).

5. Claude Code v2.1.281 (Sept 23) — Bedrock assume_role + guardrail on the Claude Apps Gateway, three security fixes

• What changed. v2.1.281 (Sept 23) shipped nine Added entries and roughly fifty Fixed entries. The week-1 enterprise additions: Bedrock assume_role on the Claude Apps Gateway; a Bedrock {guardrail: {id, version}} enforcement setting wired to the same gateway. The week-1 MCP additions: claude plugin validate now surfaces silently-dropped MCP entries in the validation report; MCP URL-mode elicitation is supported on the 2026-07-28 protocol. The three security fixes (each closing a real attack class): (i) rm -rf "$(pwd)" is now blocked in auto mode (the command substitution is a long-standing destructive-escape vector); (ii) on macOS, the /.vol/ device path was reachable through file tools that should have been pre-approval-bound to a workspace, and the pre-approval set now blocks it; (iii) NUL bytes in permission rule wildcard expansion are now blocked, so a rule like \<em>\0</em>/... cannot expand past the intended set. A prompt-cache preservation cluster also shipped: sessions resumed across a permission prompt no longer re-cache the system prompt unnecessarily.
• Who is affected. Every Claude Code operator — but the surgical cohorts are: AWS Bedrock routing teams (gain assume_role and guardrail enforcement on the Apps Gateway); Claude Code plugin authors (lose a class of silent installs); MCP server authors on the 2026-07-28 protocol (gain URL-mode elicitation); operators running Claude Code in auto mode with read-only Bash paths that include $(pwd) (lose a destructive-escape vector); macOS operators with file pre-approval lists (gain a true path-confusion guard); and anyone with permission-rule wildcards constructed from outside data (gain an injection guard).
• Cost or limit impact. None directly on pricing. The prompt-cache preservation cluster reduces unnecessary re-caching on session resumption across permission prompts, which lowers input-token cost on resume-heavy workflows; the exact cache-write savings depend on cache-write rate and prompt size, neither of which is re-measured here. Inferred: treat the cache savings as a small, plausible reduction; do not attribute a specific dollar figure.
• What could break. Plugin authors who relied on claude plugin validate silently dropping unrecognized MCP entries now see the entries surface as validation errors and must fix the marketplace entry (an actual operational fix, not just a behavior change). Auto-mode bash configurations that depended on rm -rf "$(pwd)" working now fail with a clear prompt. macOS file pre-approval lists that include /.vol/ access paths must be re-scoped. Permission rules that were constructed from untrusted input and used \0 as a delimiter now treat those rules as malformed — a safer default but a deployment one-liner change.
• What to do. Upgrade to v2.1.281 or later. Re-run claude plugin validate over your marketplace entries. Re-evaluate Bedrock routing through the Apps Gateway — set guardrails per-environment; do not set guardrails on individual sessions. Audit auto-mode bash configs for rm -rf "$(pwd)" and similar destructive substitutions. Audit macOS file pre-approval lists. Audit permission-rule construction paths for untrusted-input NUL-byte injection — particularly if any tool programmatically builds rule strings. Inferred: "Whether MCP URL-mode elicitation on 2026-07-28 protocol affects existing MCP servers transparently or requires explicit server opt-in" — not independently verified; treat as inferred.
• Primary sources. Claude Code v2.1.281 release page; anthropics/claude-code releases atom feed (v2.1.281 updated 2026-09-23T19:19:15Z).

6. Claude Code v2.1.282 (Sept 24) — managed-settings hardening closes three more attack classes

• What changed. v2.1.282 (Sept 24) shipped a managed-settings hardening sweep. Three of the closes are security-relevant: (a) macOS CLAUDE.md symlink traversal — a CLAUDE.md followed past a workspace root into a macOS symlink chain is now blocked at the managed-settings layer; (b) Boolean lock-key typo bypass — managed-settings lock keys that previously typo-matched a false value silently lost their lock (e.g., a misspelled enabled: fals defaulted to off); the parser now treats typo'd values as malformed and refuses to apply the rule; (c) plugin-manifest self-approval bypass — a plugin marketplace entry that embedded its own approval scope previously self-approved bypass of the install policy, and the managed-settings layer now reads install policy against the plugin manifest before approval is granted. The release also added store.readiness_grace_seconds for Postgres failover: a configurable grace window during which the managed store treats the database as healthy even if the readiness probe is briefly unhealthy, to avoid flapping during a brief failover.
• Who is affected. Every Claude Code operator running an enterprise deployment with managed settings. The narrowest cohorts: macOS operators where workspace roots can include symlink chains; managed-settings operators who built typed-boolean configs out of YAML or other stringly-typed inputs; plugin marketplace operators who want install-policy enforcement to be the source of truth, not self-approved by the plugin entry. Database-backed deployments gain a Postgres failover grace knob.
• Cost or limit impact. None on pricing. The store.readiness_grace_seconds knob has operational cost implications: set it too low and you flap on brief failovers; set it too high and a stuck database looks healthy for too long. The default is platform-default; operators should pick a small multiple of their expected failover time. Inferred: the specific recommended default is not independently verified for this roundup.
• What could break. macOS workspaces that previously relied on following a symlinked CLAUDE.md past the workspace root will now block. Managed-settings lock typos that previously silently disabled the lock now refuse to apply — admins deploying a typo on a release day discover it on the next push. Plugin manifests that self-approved previously may now refuse to install until the manifest is updated to be install-policy-compliant. Database HA configurations: the grace knob may need to be tuned for an existing failover SLA.
• What to do. Upgrade to v2.1.282 or later. Audit managed-settings lock keys for typo'd booleans in your YAML / JSON / TOML — particularly any stringly-typed config sources. Audit plugin marketplace entries for install-policy alignment and re-publish where the manifest self-approved before. Audit macOS workspace roots for symlink chains that previously crossed into the workspace. Pick a store.readiness_grace_seconds that is a small multiple of your Postgres failover SLA — and test it. Inferred: tuning defaults are derived; treat as inferred.
• Primary sources. Claude Code v2.1.282 release page (cross-referenced via the change-intelligence coverage of the same release at https://mr.technology/payloads/claude-code-2-1-282-managed-settings-hardening-macos-symlink-plugin-self-approval-sep-2026).

7. OpenClaw v2026.9.5 (Sept 19) atomic updates, plugin hot reload, conversation sharing, GPT-Live — then v2026.9.6 macOS DMG rebuild four hours later (Sept 24)

• What changed. OpenClaw shipped two stable-channel releases a week apart. v2026.9.5 (Sept 19) introduced atomic updates: when openclaw update lands a broken build, the Gateway rolls back to the previous known-good build before returning control to the operator, and the half-state path is closed. It also added plugin hot-reload (so plugin edits are picked up without a Gateway restart), conversation sharing (so a user can hand a session to another user without exporting and re-importing), and GPT-Live integration (so the OpenClaw runtime can route through the OpenAI GPT-Live family where supported). v2026.9.6 (Sept 24) rebuilt the macOS DMG four hours after the original 2026.9.6 macOS build crashed at launch on macOS Gatekeeper (#156861 → fix #156881, rebuilt notarized DMG shipped at 09:52 UTC). The npm package was unchanged. The headline stats for the v2026.9.6 merge window: 178 direct commits, 2,614 pull requests, 351 contributors. Inferred: the macOS DMG rebuild is a release-quality event, not a routine one; treat the four-hour gap as a primary-source figure from the v2026.9.6 release page.
• Who is affected. Every OpenClaw operator — but the surgical cohorts are: operators who had the original 2026.9.6 macOS DMG crash on launch (they need to install the rebuilt DMG once); operators who have been bitten by failed Gateway updates and want the rollback behavior (gain v2026.9.5); operators on the GPT-Live family (gain a routing path); operators handing sessions between users (gain conversation sharing).
• Cost or limit impact. None on pricing. The atomic-update rollback reduces the operational cost of failed updates by closing the manual-recovery loop. macOS operators who did not install the original DMG are unaffected; operators who installed it and got the launch crash need one reinstall (no data loss, since the npm path was unchanged).
• What could break. Plugin authors whose plugins depended on a Gateway restart for code reloading now see hot-reload — test that before relying on it. macOS operators who installed the original 2026.9.6 DMG before 09:52 UTC and whose Gatekeeper blocked the launch may still have a half-state local install; reinstalling once from the rebuilt DMG clears it. Conversation-sharing flows that previously assumed exports and re-imports will need to verify the new sharing path. GPT-Live routing now spans two providers' governance surfaces — keep an eye on the OpenClaw-side audit.
• What to do. Upgrade to v2026.9.5 or later on every host. If you are on macOS and installed the original 2026.9.6 DMG before the rebuild, install the rebuilt DMG once. Re-validate plugin hot-reload on a representative plugin before relying on it. Test the conversation-sharing path on a representative handoff before relying on it. Verify GPT-Live routing on a representative workflow before scaling. Inferred: "whether the rebuilt DMG's notarization held on macOS Gatekeeper" — not independently verified for this roundup.
• Primary sources. OpenClaw v2026.9.5 release page; OpenClaw v2026.9.6 release page (verified 2026-09-24 20:08 UTC); OpenClaw releases atom feed.

Sidebar A: the build pipeline recovered (Sept 25, 15:04 UTC)

Between Sept 21 and Sept 25, the auto-rebuild pipeline was recurrently non-recoverable: the Next.js 16 webpack stage (isolated next build --webpack try 1/3) wedged at "Creating an optimized production build ..." with no completion entry, the spawn PID died, the BUILD_ID did not advance, and newly authored .tsx pages returned HTTP 404. The same defect class — a webpack WasmHash transient race (TypeError: Cannot read properties of null (reading 'length') at WasmHash._updateWithBuffer) and a related silent-hang variant — surfaced in INFRA_ISSUE logs across multiple publications in the same window (Sept 21, Sept 22, two on Sept 24, one on Sept 25 morning). The Sept 25 evening sweep landed at the start of a recovered window: BUILD_ID advanced to Lf5-eDUuD03czyUlG-BnY at 15:04 UTC, and the payload routes for the recently authored articles returned HTTP 200 locally. Inferred: the recovery is reproducible but not guaranteed — three further rebuild attempts were spawned on Sept 25 afternoon before that one succeeded, each dying at the same step. Treat the pipeline as healthy-but-fragile, not as restored.

• Primary sources. /var/log/auto-rebuild.log — Sept 21–25 entries; frontend/.next/BUILD_ID mtime 2026-09-25 15:04; local curl http://localhost:3000/payloads/<slug> for the affected slugs.

Sidebar B: six Claude Code versions in eight days

Between Sept 18 and Sept 26, Claude Code shipped v2.1.277, v2.1.278, v2.1.280, v2.1.281, v2.1.282, and v2.1.283 prep. Not all of them are security drops — v2.1.277 added AGENTS.md discovery when CLAUDE.md is absent and tightened the Gateway egress boundary; v2.1.278 moved auto-mode to server-side classifier with no separate auto-mode charge on the client side; v2.1.280 made Opus 5.5 the default; v2.1.281 added Bedrock assume_role + guardrail on the Apps Gateway plus three security fixes; v2.1.282 closed three more attack classes in managed settings. The point: the release cadence is now weekly-or-better, and two consecutive weeks have shipped version pairs (v2.1.275 + v2.1.277; v2.1.281 + v2.1.282) of closely-spaced security sweeps. If you are managing a Claude Code fleet, the operational question is not whether to upgrade but how to keep the upgrade cadence sustainable — and which subset of releases you put behind a soak test before promoting to a fleet.

• Primary sources. github.com/anthropics/claude-code releases — v2.1.277 through v2.1.282.

What we deliberately did not include

• Routine funding announcements, vague roadmap posts, and recycled launch stories. The bar is primary-source change with practical consequences.
• Pre-release alpha/beta churn from MCP TypeScript SDK, Gemini CLI, and similar. The desk tracks these as NOVELTY_LOW source-change items and surfaces them when they ship stable.
• Cursor Rollouts + Security Review entry (Teams / Enterprise gating, changelog date ambiguity). Held below standalone-news threshold for the roundup; flagged for next week's morning desk review.
• Firsthand claims we did not make. Every item above is sourced from the upstream release notes, changelog, release page, or status page. Where a claim is inferred (cost percentages, error code specifics, MCP URL-mode elicitation opt-in semantics, store.readiness_grace_seconds tuning defaults), it is labeled.

What to watch next week

• Opus 5.5 migration fallout. Pin your Claude Code default and audit your request shapes before team members hit Enter on a new session.
• Refusal-billing scale-up. Anthropic characterised the three billed categories as "low volumes of false positives" — verify on your own eval pipeline.
• Claude API incident cadence. Third incident in twelve days. Audit your fallback tier chains and wire the Status-page webhook if you have not already.
• OpenAI Codex rust-v0.158.0+. The v0.157.0 hardening closed three real gaps; expect the next drop to add to the enterprise trust surface.
• OpenClaw v2026.9.7 / v2026.7.36 stable cadence. Watch the next patch wave for notarization follow-up and any rollout of the conversation-sharing surface to additional providers.
• Claude Code v2.1.283+. Two consecutive weeks of paired security sweeps; the cadence suggests another patch wave before next Friday's roundup.
• The build pipeline. Treat as healthy-but-fragile; the Sept 25 recovery was preceded by three failed attempts.

Originally published: 2026-09-26 02:00 Berlin / 2026-09-26 00:00 UTC Last verified: 2026-09-26 00:00 UTC No corrections at this time.

— Mr. Technology