Dateline: October 16, 2026 | Next update: October 23, 2026
A quieter week than the last three, and more useful for it. The headline is Gemini 4 Argon reaching general availability — along with documentation that quietly corrects the most-quoted number from its announcement. Underneath, Anthropic spent the week on observability: tracing, fast mode graduating, and a cost feature that works while you are not at the desk. OpenAI published exactly one changelog entry.
Two dates to put in the calendar before the next issue: October 19, when GitHub Copilot retires six models announced a month ago, and October 23, when Google shuts down deep-research-pro-preview-12-2025.
Claude / Anthropic
★ OpenTelemetry export reaches beta for the API and Managed Agents
Messages API requests and Managed Agents sessions can now be traced in your own observability backend. The Claude SDK emits spans in Python, TypeScript, C#, Go, Java, PHP and Ruby; to get API spans you add an OpenTelemetry destination in the Claude Console.
This closes a gap that gets painful at exactly the wrong moment. An agent that takes ninety seconds is a support ticket you cannot answer without knowing which call was slow — the model, a tool, or your own code between them. Spans in the same backend as the rest of your stack mean the agent stops being the one component you reason about from logs. Seven languages at beta is also broad enough that most teams will not be waiting for their SDK.
Status: beta | Traced: Messages API requests and Claude Managed Agents sessions | SDK span emission: Python, TypeScript, C#, Go, Java, PHP, Ruby | API spans: add an OpenTelemetry destination in the Claude Console | Also October 14: Managed Agents scheduled deployments now available on Claude Platform on AWS | Related, in Claude Code 2.1.300–2.1.301: `cache_creation_5m_tokens` and `cache_creation_1h_tokens` split cache writes by lifetime in the `api_request` event, and `classifier_verdict` and `permission_mode` were added to `tool_decision` events
Best for: Anyone running agents in production who has been diagnosing latency from log lines. Wire the destination before you need it, not during the incident.
★ Fast mode for Opus 5.5 leaves research preview
Fast mode on Opus 5.5 no longer needs an access request. The fast-mode-2026-02-01 beta header is no longer required, though requests that still send it keep working, and Opus 5.5 now carries a default fast mode rate limit in the Start, Build and Scale tiers. It is also available for Claude in Microsoft Foundry on Azure-hosted deployments.
Read the second half of the note carefully: fast mode for Claude Opus 5 and Claude Opus 4.8 is still in research preview, and those requests must keep sending the beta header. If you have a codebase that sets the header globally and routes across several Opus versions, removing it will break the older two while leaving 5.5 working — the kind of partial failure that takes a while to attribute.
Model: `claude-opus-5-5` | Access: no request needed | Beta header: `fast-mode-2026-02-01` no longer required; requests still sending it keep working | Rate limits: default fast mode rate limit in the Start, Build and Scale tiers | Platforms: also on Claude in Microsoft Foundry for Azure-hosted deployments | Still in research preview: fast mode for Claude Opus 5 and Claude Opus 4.8 — keep sending `fast-mode-2026-02-01` on those
Best for: Interactive workloads on Opus 5.5 that were waiting on preview access. Leave the header in place if you also call Opus 5 or 4.8.
★ Claude Code compacts idle conversations before the cache expires
A conversation over 200,000 tokens that sits idle for roughly 54 minutes is now compacted automatically — just before its one-hour prompt cache expires.
The timing is the whole idea. A long conversation whose cache has lapsed is expensive to resume: the next message re-sends the full prompt at the uncached rate. Compacting at 54 minutes means the session comes back smaller instead of coming back costly, and it happens while nobody is waiting. Both numbers are adjustable — `CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS` changes the 200K threshold, and `"idleCompaction": false` in settings.json turns it off entirely.
Version: 2.1.299 (published October 14, 2026) | Trigger: conversation over 200K tokens idle for about 54 minutes, before the one-hour prompt cache expires | Threshold: `CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS` | Disable: `"idleCompaction": false` in settings.json | Also in 2.1.299: `CLAUDE_CODE_DISABLE_FILE_MENTIONS` gives Claude the paths of `@` file mentions to read with its file tools rather than inlining the file contents | Also in 2.1.296: `autoCompactWindow` in subagent frontmatter, so a subagent can auto-compact earlier than the main conversation
Best for: Long analysis sessions picked up across a working day. If your team leaves big conversations open over lunch, this is a quiet reduction in the bill.
★ The Compliance API returns the instructions people set for themselves
A new endpoint returns the text each member saved as Instructions for Claude in claude.ai, and as Global instructions in Cowork.
Custom instructions are a blind spot in most governance reviews. They are written by individuals, never reviewed, and they change how the model behaves for that person on every conversation — which is exactly the sort of thing a compliance team is supposed to be able to see. Being able to read them organisation-wide turns an unknown into an auditable field. It is also, plainly, a privacy-sensitive capability: it should sit behind the same approval as any other bulk read of what employees have written.
Endpoint: `GET /v1/compliance/organizations/{org_uuid}/user_preferences` | Authentication: Compliance Access Key | Scope: `read:compliance_user_data` | Returns: each member's "Instructions for Claude" from claude.ai and "Global instructions" from Cowork | Status: beta, Claude Enterprise
Best for: Compliance teams who have been asked what employees actually told the model to do. Decide who may call this before you turn the key on.
Plans and pricing
No price changes were published this week. The commercial change is access rather than rate: fast mode on Opus 5.5 no longer requires a preview request, and now carries a default rate limit by tier.
No published Claude API price changes between October 9 and 16, 2026 | Opus 5.5: $4/$20 per MTok, cache read $0.20 | Sonnet 5.5: $2/$10 per MTok, cache read $0.10 | Haiku 5.5: $0.10/$0.50 per MTok up to 100K-token prompts, $0.50/$2.50 above | Fable 5.1: $10/$50 per MTok | Fast mode on Opus 5.5: default rate limit in the Start, Build and Scale tiers | No model retirements announced this period
ChatGPT / OpenAI
★ Decisions runs natively inside GPT-Live
The Decisions API, which entered beta the week before, is now wired directly into GPT-Live. An application can act while the user is still speaking, or react to what Live says, with Live calling Decisions through native delegation — or decisions triggered as input and output transcript deltas arrive.
Acting mid-sentence is a different product than acting after a turn ends. A voice agent that waits for the caller to finish, classifies, then routes, has a pause in it that everyone on the line can hear. One that classifies on transcript deltas can have the right screen open before the sentence lands. That is the gap between a demo and something a contact centre would actually put on a queue — and it is the first concrete payoff from the Decisions endpoint being narrow and fast rather than general.
Models: `gpt-live-1` with `gpt-6-luna` | Endpoint: `v1/live/sessions` | Mechanisms: native delegation, where Live calls Decisions directly; or decisions triggered as input and output transcript deltas arrive | Capability: act while the user is still speaking, or react to what Live says | Pricing and service tiers: none listed with the release
Best for: Voice workflows where the dead air after a caller stops talking is the thing customers complain about.
A single entry, and nothing else
Other than the GPT-Live integration, OpenAI published nothing to its developer changelog in this window. The previous entry was Ultrafast for GPT-6.1 Sol on October 8, which we covered last week.
Two weeks after DevDay this is unsurprising — most of what was announced on September 29 is still rolling out rather than being added to. Several of those items were dated "coming weeks" at announcement, including collaborative slides and GPT-6.1 Sol Ultrafast, so expect those to appear as changelog entries rather than as new announcements.
Plans and pricing
No pricing or tier changes were published in this window.
No published OpenAI price changes between October 9 and 16, 2026 | In effect from October 6: usage tiers reduced from five to three (Build, Launch, Grow) | GPT-6.1 Sol: $2 input / $0.10 cached input / $2.50 cache write / $10 output per 1M tokens, up to 272K input | Decisions API: beta, pricing on the pricing page
Gemini / Google
★ Gemini 4 Argon reaches general availability — and the output number is corrected
Two weeks after launching to trusted cyber defenders through the Fairwind Program, Argon is generally available in Google AI Studio and the API, described as Google's most capable model for complex coding, agentic, mathematics and science work. It takes text, image, video, audio and PDF in, and returns text, with Batch, Flex and Priority inference all supported.
The documentation corrects the headline figure from the announcement. Argon's context window is 1,048,576 tokens shared by input and output, and the default maximum output for a synchronous request is 65,536 tokens — the million-token output is available only with long decode. We reported the million-token output limit on September 30 as Google announced it; the shipped default is sixteen times smaller. If you planned a single-pass job around a million tokens of output, that plan needs the long-decode path, not the default one.
Model: `gemini-4-argon`, listed as Stable | Context: 1,048,576 tokens, shared by input and output | Max output: 65,536 tokens by default for synchronous requests; up to 1,048,576 with long decode | Input: text, image, video, audio, PDF | Output: text | Consumption: Batch API, Flex inference, Priority inference | Availability: not on the Free tier or the Google AI Pro plan | Also added to GitHub Copilot on October 13, the same day
Best for: Complex coding and agentic work, now that access is open. Check whether your use case needs long decode before costing it.
⚠ Argon's introductory price has an end date: November 30
Argon runs at $2 per million input tokens and $10 per million output through November 30. From December 1 those become $4 and $20 — a doubling on both sides.
Six weeks of half-price is long enough to build something on and short enough to forget. Cached input doubles too, $0.10 to $0.20, and it is worth knowing that Argon supports implicit caching only — explicit caching is not available, so you cannot pin a prefix the way you might elsewhere. Batch runs at half the standard rate throughout: $1/$5 now, $2/$10 from December.
Standard, paid tier, per 1M tokens — through November 30, 2026: $2.00 input, $10.00 output (including thinking), $0.10 cached input | From December 1, 2026: $4.00 input, $20.00 output, $0.20 cached input | Caching: implicit only; explicit caching not supported | No prompt-length tiers listed for Argon | Batch: half the standard rate — $1.00/$5.00 now, $2.00/$10.00 from December 1 | Not available on the Free tier or the Google AI Pro plan | For comparison, `gemini-3.8-flash`: $0.75/$3.75 through December 31, 2026, then $1.50/$7.50
Best for: Put December 1 in the same calendar entry as the pilot. A proof of concept costed in October reads very differently in December.
⚠ Deadline: the old Deep Research agent shuts down on October 23
A reminder rather than news. deep-research-pro-preview-12-2025 stops working one week from this issue.
Shutdown: October 23, 2026 | Migration: change the `agent` parameter in `interactions.create` to `deep-research-preview-04-2026` or `deep-research-max-preview-04-2026` | Also deprecated October 8, with requests auto-routed and no shutdown date: `gemini-3.7-flash` → `gemini-3.8-flash`, `gemini-3.5-flash` → `gemini-3.6-flash` | Also deprecated October 6, no shutdown date: `gemini-3.1-flash-image` → `gemini-nano-banana-2.1`
Best for: If this is still in a pipeline somewhere, it is a one-parameter change and it is due this week.
Plans and pricing
Argon's rate card is the only pricing movement, and it is dated.
`gemini-4-argon`: $2/$10 per MTok through November 30, 2026; $4/$20 from December 1 | Cached input: $0.10 then $0.20, implicit caching only | Batch: $1/$5 then $2/$10 | `gemini-3.8-flash`: $0.75/$3.75 through December 31, 2026; $1.50/$7.50 from January 1, 2027; cached input $0.075 then $0.15 | Context caching storage: $0.50 per 1M tokens per hour, $1.00 from January 1, 2027 | Free tier: no charge for input or output, and no Argon access
Microsoft Copilot
★ Agentic autofix reaches general availability
Autofix for code scanning alerts now works the way a developer would: it explores the relevant files, proposes a fix, reruns CodeQL to confirm the alert is closed, and iterates if it is not. It then opens a draft pull request with a summary of the fix, why it closes the alert, and the validation steps it took.
Rerunning CodeQL to verify is the part that matters. A suggested fix that has not been checked is a suggestion; one that has closed the alert and shown its working is a reviewable change. It can also batch up to 25 alerts into a single pull request, which turns a backlog of similar findings into one review rather than twenty-five. Where it replaces the free Generate Fix option, the alert now shows an Assign to Copilot button. It only works on alerts from the repository's default branch.
Behaviour: explores relevant files, proposes a fix, reruns CodeQL to verify the alert closes, iterates if needed, opens a draft pull request with summary, rationale and validation steps | Batch: up to 25 alerts in one pull request | Trigger: UI ("Assign to Copilot") or the REST API | Limitation: default-branch alerts only | Plans: GitHub Code Security and GitHub Advanced Security on GitHub Enterprise Cloud, including the data residency variant | Enablement: an enterprise administrator enables the "Agentic autofix" policy in Advanced Security policies, then an organization or repository administrator turns it on in Code Security settings; a Copilot billing method must be defined | Cost: consumes AI Credits from the Code Security budget, plus GitHub Actions minutes; usage appears under "AI usage" → "Advanced Security" in enterprise billing, and SKU-level budgets can be set
Best for: Security backlogs where the same class of finding repeats. Set the Code Security AI credit budget before enabling it, not after the first month's bill.
⚠ Three more models retire on November 16
The third deprecation list in a month, each entry with a named replacement.
Claude Opus 4.8 → Claude Opus 5.5. Claude Opus 5 → Claude Opus 5.5. GPT-5.6 Terra → GPT-6.1 Sol or GPT-6 Luna. One exception is worth noting: Claude Opus 4.8 stays available to individual Copilot subscribers on annual plans, so they keep an Opus option. Under default model enablement, Enterprise and Business customers get the replacements automatically — unless an administrator turned the global default off or explicitly disabled a model, in which case the replacements must be enabled through model policies.
Retirement date: November 16, 2026 | Claude Opus 4.8 → Claude Opus 5.5 (remains available to individual subscribers on annual plans) | Claude Opus 5 → Claude Opus 5.5 | GPT-5.6 Terra → GPT-6.1 Sol or GPT-6 Luna | Default model enablement: replacements activate automatically for Enterprise and Business | If the global default is off: enable replacements through model policies in Copilot settings | No action needed to remove deprecated models after retirement | Still pending from the September 18 notice: six models retire October 19, 2026
Best for: Whoever owns the model policies. The October 19 list lands first — three days from this issue.
★ Gemini 4 Argon lands in Copilot on the day it reaches GA
Gemini 4 Argon appeared in Copilot on October 13, the same day Google made it generally available — the fourth consecutive same-day model addition.
Gemini 4 Argon in GitHub Copilot: October 13, 2026 | Also this week: multi-repo support and other improvements for Copilot in Slack and Microsoft Teams (October 15); HydraFusion updates and a new Max option (October 14); enterprise managed settings enforcement updates (October 13); client telemetry gains user attribution and configurable capture (October 12); new controls and chat improvements in Copilot for JetBrains (October 10)
Best for: Teams who evaluate models inside Copilot rather than against each vendor's API. Same-day availability is now the pattern across all four labs.
Plans and pricing
No seat price changes. The cost movement is in what draws on AI Credits, which now includes agentic autofix at general availability.
No Copilot seat price changes announced between October 9 and 16, 2026 | Agentic autofix: consumes AI Credits from the Code Security budget plus GitHub Actions minutes; SKU-level budgets available | Model added: Gemini 4 Argon, October 13 | Retirements: six models on October 19, 2026 (announced September 18); three models on November 16, 2026 (announced October 15) | Client telemetry: user attribution and configurable capture, October 12
Best for: Set the Code Security AI credit budget this week. Agentic autofix going GA is the first Copilot security feature that will show up on a bill at scale.