Dateline: September 25, 2026 | Next update: October 2, 2026
A week with one very large story and one conspicuous silence. Anthropic shipped Claude Opus 5.5 and made it the default across every paid plan, cutting the input price of its mid-tier frontier model in half. Google moved its 3.8-generation speech models to general availability and began steering new projects away from the 2.5 generation. GitHub Copilot gave six models a retirement date and a named replacement each. OpenAI published nothing to its developer changelog all week — not a quiet quarter, but a deliberate hold: DevDay 2026 was four days after this window closed.
Claude / Anthropic
★ Claude Opus 5.5 ships, and becomes the default everywhere
Anthropic now points at Opus 5.5 as the starting point for most workloads, a recommendation that until this week belonged to whichever Sonnet was current. It carries a one-million-token context window, 128K maximum output, and adaptive thinking that is always on, with the effort parameter defaulting to medium.
The price is the part worth reading twice. Opus 5.5 runs at $4 per million input tokens and $20 per million output — against $5/$25 for the Opus 5 it supersedes. A frontier-tier model got cheaper rather than more expensive, and prompt cache reads sit at 5% of base input rather than the usual 10%, which works out to $0.20 per MTok.
The default changed as well as the lineup. In Claude Code 2.1.280, the default model on Pro and Team Standard moved from Sonnet to Opus, matching what Max, Team Premium and Enterprise already did. If your team is on Pro and has never touched the model picker, the model answering them changed this week without anyone choosing it. That is worth knowing before the first invoice, not after — though with Opus 5.5 priced below Opus 5, the direction of the surprise is unusually friendly.
Model ID: `claude-opus-5-5` (alias and pinned snapshot are the same string) | Context: 1M tokens | Max output: 128K synchronous, up to 300K on the Batches API with the `output-300k-2026-03-24` beta header | Pricing: $4 input / $20 output per MTok; cache reads at 5% of base input ($0.20/MTok); Batch API 50% off | Thinking: adaptive, always on | Default effort: `medium` | Reliable knowledge cutoff: June 2026 | Retirement: not sooner than September 22, 2027 | Bedrock: `anthropic.claude-opus-5-5` | Google Cloud and Microsoft Foundry: `claude-opus-5-5`
Best for: Long-running agentic coding and knowledge work. If you moved workloads off Opus 5 to control cost, the arithmetic that pushed you away has changed — re-run it before renewing anything.
★ Claude Code reads AGENTS.md when there is no CLAUDE.md
A small change with a diplomatic purpose. AGENTS.md is the convention several other coding agents settled on; CLAUDE.md is Anthropic's. From 2.1.277, a project that has no CLAUDE.md gets read from AGENTS.md instead, which is selectable under "Project instructions" in /config.
For a team running more than one coding agent across the same repository, this removes a duplicated file that drifted the moment someone edited one and not the other. The precedence is worth remembering: `CLAUDE.md` still wins where both exist, so adding `AGENTS.md` to a repo that already has `CLAUDE.md` changes nothing at all.
Version: 2.1.277 (published September 18, 2026) | Behaviour: `AGENTS.md` is read only in a project with no `CLAUDE.md` | Configuration: "Project instructions" in `/config` | Precedence: `CLAUDE.md` takes priority where both files exist
Best for: Repositories where more than one agent tool is in use, and the instructions file has been maintained twice.
★ Administrators can pin an exact model and block specific ones
Two managed settings landed for organisations that need model choice to be a decision rather than a default. availableModelsMatch set to "exact" makes an availableModels entry allow only the model version it names, so a newly released model stays blocked until someone adds it to the list. deniedModels blocks named models outright, even when availableModels would otherwise permit them.
The timing is not a coincidence. In the same week that a new default Opus rolled out to Pro and Team Standard, Anthropic shipped the control that stops exactly that from happening unannounced. Organisations with model governance requirements — regulated industries, anyone whose evals are pinned to a specific version — can now make new releases opt-in rather than opt-out.
Version: 2.1.283 (published September 25, 2026) | `availableModelsMatch: "exact"` — an `availableModels` entry allows only the version it names | `deniedModels` — blocks specific models even when `availableModels` allows them | Also in 2.1.283: `/doctor prompt-audit` (alias `/checkup prompt-audit`) audits CLAUDE.md files, skills, agents and commands for prompting patterns written for older models
Best for: Platform teams who need a model upgrade to be a change they approve, not one they discover.
★ Auto mode's classifier moves to the server, and stops being billable
Auto mode decides which permission prompts it can handle without asking you. That decision used to be made by a model call you paid for. From 2.1.278 it defaults to a server-side classifier, which does not charge for classification.
A `/status` row named "Auto mode server" now tells you which classifier this session is using, so the change is visible rather than assumed. Two versions later, in 2.1.282, the server-side classifier also became the default on a direct Anthropic API connection when telemetry is off — with `CLAUDE_CODE_AUTO_MODE_SERVER=0` as the opt-out.
Version: 2.1.278 (September 19) and 2.1.282 (September 24) | Default: server-side classifier for Claude API and Enterprise users, and on Bedrock, Vertex, Foundry and gateways | Billing: classification is not charged | Visibility: "Auto mode server" row in `/status` | Opt-out: `CLAUDE_CODE_AUTO_MODE_SERVER=0` | Also in 2.1.281: `/insights` now estimates how many permission prompts auto mode could have handled in your recent sessions
Best for: Teams who left auto mode off because the classification calls were an unpredictable line item.
Plans and pricing
The only pricing movement was downward. Opus 5.5 enters at $4/$20 per MTok against Opus 5's $5/$25, and Opus 5 remains available for anyone who has pinned to it. Claude Fable 5.1 is unchanged at $10/$50. No model retirements were announced in this window.
Opus 5.5: $4/$20 per MTok, cache reads $0.20/MTok (5% of base input) | Opus 5: $5/$25 per MTok, still available | Fable 5.1: $10/$50 per MTok | Haiku 4.5: $1/$5 per MTok | Batch API: 50% off base price on all models | Plan defaults: Pro and Team Standard moved from Sonnet to Opus on September 22 | No retirements announced this period
Best for: Anyone whose cost model was built on Opus 5 pricing. The input side is 20% cheaper and cache reads are half what the standard rate would be.
ChatGPT / OpenAI
A week with nothing shipped, by design
OpenAI published no developer changelog entries at all during this window. The last one before it was API key creation governance on September 15, which we covered in the previous issue. The next came on September 29 — four days after this window closed — when DevDay 2026 delivered more than twenty announcements at once.
This is a reporting boundary, not a quiet quarter. Everything OpenAI had been holding landed on September 29: GPT-6.1 Sol, computer use in the Agents API, and an ultrafast service tier among them. All of it sits outside the 18–25 September window and belongs to the next issue, where we will cover it from the published documentation rather than from conference coverage.
OpenAI Academy expands into role-specific learning paths
OpenAI added courses aimed at particular roles rather than at ChatGPT in general — developers, leaders, educators and college students — joining the existing Apply AI at Work course. The developer track, Build with AI, is aimed at teams using Codex or building on the OpenAI API, and covers solution design, evaluations, agents, retrieval and operating AI systems in production.
Course assessments now carry a badge on completion. For a company trying to move a whole team from ad-hoc prompting to something repeatable, a shared course with an assessment at the end is a more practical starting point than a policy document nobody reads. OpenAI put the programme's two-year totals at over 250 events and more than four million people engaged.
New pathways: AI for College Students; Build with AI (developers and technical teams); plus tracks for leaders and educators | Existing: Apply AI at Work | Build with AI covers: planning and implementing changes across the software development lifecycle, solution design, evaluations, agents, retrieval, production operation | Recognition: OpenAI Academy course badge on passing the assessment
Best for: Operations leads who need a shared baseline across a team before rolling out agents, and who would rather not write the training themselves.
An advisory group on mathematics and AI
OpenAI announced an advisory group on mathematics and artificial intelligence. No product change attaches to it in this window.
Best for: Context rather than action. Noted here for completeness — it is one of only two OpenAI announcements in the period.
Gemini / Google
★ Gemini 3.8 Flash TTS and Flash-Lite TTS reach general availability
Google's text-to-speech models left preview in two tiers. gemini-3.8-flash-tts is the flagship, described by Google in terms of studio-grade voice fidelity, nuanced acting, regional dialects and stability across long multi-turn output. gemini-3.8-flash-lite-tts is the cost-efficient tier and replaces gemini-3.1-flash-tts-preview.
Three capabilities arrive with GA: voice design, voice replication gated behind consent verification, and a library of more than 150 prebuilt and custom voices. The consent gate on replication is the detail worth noting — a vendor putting a verification step in front of voice cloning is making a product decision that will show up in procurement questionnaires, and it is easier to answer a compliance question about a feature that was built with the gate than one where you added it yourself.
Flagship: `gemini-3.8-flash-tts` — studio-grade fidelity, nuanced delivery, regional dialects, long-form multi-turn stability | Cost-efficient: `gemini-3.8-flash-lite-tts` — replaces `gemini-3.1-flash-tts-preview` | Features at GA: voice design; voice replication with consent verification; 150+ prebuilt and custom voices | Status: generally available on the Gemini API
Best for: Anyone shipping voice output who has been waiting for a preview model to stabilise before committing. Long-form stability is the specific claim to test against your own scripts.
⚠ Gemini 2.5 access narrows to teams already using it
Access to the 2.5 generation is now limited to users who have actively used those models before. They remain available through the API and Google is explicit that they are not deprecated — but new projects are steered to 3.5 Flash-Lite or 3.8 Flash instead, which Google frames as preserving capacity for both legacy and new applications.
Read the mechanism rather than the label. "Restricted, not deprecated" means existing integrations keep working and no retirement date has been set, but a new project — a new Google Cloud project, a new environment, a fresh proof of concept — cannot reach for a 2.5 model. If your production stack is pinned to 2.5 and your staging environment is rebuilt from scratch, that is the failure mode to check for before it surprises you.
Effective: September 18, 2026 | Restriction: access limited to users who have previously used the models actively | Availability: still served through the API | Deprecation: none announced — Google states the models are not deprecated | Recommended for new projects: Gemini 3.5 Flash-Lite or Gemini 3.8 Flash | Google's stated reason: maintaining capacity for both legacy and new applications
Best for: Teams still on 2.5 in production. Nothing breaks today, but the path to 3.8 Flash just became the only path for anything new.
Plans and pricing
No pricing changes were published for the Gemini API in this window. The TTS models moved from preview to general availability, which changes their support commitment rather than their rate card.
No published Gemini API price changes between September 18 and 25, 2026 | `gemini-3.8-flash-tts` and `gemini-3.8-flash-lite-tts`: preview to GA | `gemini-3.1-flash-tts-preview`: superseded by the Flash-Lite TTS model | Gemini 2.5 family: access restricted, no retirement date announced
Microsoft Copilot
⚠ Six models retire on October 19, each with a named replacement
GitHub published the list and the date together, which makes this an unusually easy deprecation to plan around.
Gemini 3.7 Flash → Gemini 3.8 Flash. GPT-5.5 → GPT-5.6 Sol. GPT-5.4 → GPT-5.6 Sol. GPT-5.4 mini → GPT-5.6 Luna. GPT-5 mini → GPT-5.6 Luna. Grok 4.5 → Grok 4.6. For Enterprise and Business customers with default model enablement switched on, the replacements activate automatically — unless an administrator has turned off global defaults or disabled a specific model, in which case nothing activates and the workflow simply loses its model.
Date: October 19, 2026 | Deprecated and replacement: Gemini 3.7 Flash → Gemini 3.8 Flash; GPT-5.5 → GPT-5.6 Sol; GPT-5.4 → GPT-5.6 Sol; GPT-5.4 mini → GPT-5.6 Luna; GPT-5 mini → GPT-5.6 Luna; Grok 4.5 → Grok 4.6 | Enterprise and Business with default model enablement: replacements activate automatically | Exception: administrators who disabled global defaults or specific models must act manually | Guidance: update workflows and integrations before the date
Best for: Anyone with a model name hard-coded in a workflow file or an integration. The automatic path only covers organisations that left the defaults alone.
★ Agentic autofix now reads and writes Copilot Memory
Agentic autofix, which proposes fixes for security alerts, now consults stored memories for context before proposing one, and saves successful fix patterns back as new memories.
The loop is the interesting part rather than either half of it. A fix that worked in your repository becomes context for the next alert, and GitHub says those memories also reach Copilot code review and the cloud agent — so a pattern learned while fixing a vulnerability can shape an unrelated review later. Both agentic autofix and Copilot Memory are in public preview, so treat the behaviour as subject to change and worth watching rather than trusting unattended.
Status: public preview (both agentic autofix and Copilot Memory) | Requirement: Copilot Memory enabled for the customer | Behaviour: reviews stored memories for context when fixing an alert; saves successful fix patterns as new memories | Scope of reuse: memories also inform Copilot code review and cloud agent features
Best for: Security teams with a recurring class of alert and a house pattern for fixing it — this is where a memory earns its keep.
★ The usage metrics API breaks pull request review into three stages
A pull_request_review_times array joins the usage metrics API, splitting review latency into ready-to-first-review, first-to-final-review, and final-review-to-merge — each as a median and a 90th percentile.
Three stages instead of one number is the difference between knowing that review is slow and knowing where. A long ready-to-first-review gap is a queueing problem and usually an ownership question. A long final-review-to-merge gap is a process problem — approvals sitting unmerged — and has nothing to do with reviewer capacity. The array also carries `authored_by`, `reviewed_by` and `total_merged`, so the counts sit alongside the timings.
New array: `pull_request_review_times` | Fields: `median_minutes_ready_to_first_review`, `p90_minutes_ready_to_first_review`, `median_minutes_first_to_final_review`, `p90_minutes_first_to_final_review`, `median_minutes_final_review_to_merge`, `p90_minutes_final_review_to_merge`, `authored_by`, `reviewed_by`, `total_merged` | Present in: enterprise and organization `repos-1-day` reports
Best for: Engineering managers being asked to justify an AI spend with a cycle-time number. This is the first Copilot metric that isolates which part of review actually moved.
★ A validator for enterprise managed settings
Copilot's enterprise configuration now validates itself in-product. The checker reads copilot/managed-settings.json and copilot/team-mappings.json and reports malformed JSON, unsupported configurations and invalid team mappings, naming the affected file and the JSON path for each issue.
Silent failure is the specific problem here. A policy file with a typo does not announce itself — it simply does not enforce, and the first evidence is a developer doing something the policy was meant to prevent. Pointing at the file and the JSON path turns that into a two-minute fix.
Files checked: `copilot/managed-settings.json`, `copilot/team-mappings.json` | Detects: malformed JSON, unsupported configurations, invalid team mappings, and other errors that prevent policy enforcement | Reporting: each issue names the affected file and JSON path | Location: "Copilot settings validation" on the enterprise AI controls page
Best for: Anyone who has written a Copilot policy file and never had confirmation it took effect.
Plans and pricing
No price changes were announced. The week's commercial news is the October 19 retirement list, which changes which models your seats can reach rather than what they cost.
No Copilot price changes announced between September 18 and 25, 2026 | Model retirements: six models on October 19, 2026, replacements named | Also shipped September 18: an improved Copilot code review experience | Also shipped September 25: updates to Copilot for Slack and Microsoft Teams | Weekly release posts: September 14 batch published September 18; September 21 batch published September 25