Dateline: August 7, 2026 | Next update: August 14, 2026
No new model this week, but three substantial enterprise and infrastructure stories. On August 5, Anthropic launched inference hooks — server-side inline data loss prevention for Claude Enterprise — and on August 6 opened the public beta of self-hosted environments for Claude Code, letting enterprise teams run agent sessions on their own compute. Anthropic also confirmed it is building an in-house silicon team. Separately, the Fable 5 subscription structure finally stabilised: Max and Team Premium get permanent included access at 50% of weekly limits from July 20; Pro and Team Standard received a one-time $100 credit and moved to usage-credit billing. Claude Code's 50% higher weekly limits run through August 19. Opus 4.1 retired August 5. The Workbench and experimental prompt APIs retire in ten days (August 17). On the OpenAI side, Codex CLI 0.147.0 landed with portable Agent Plugins, persistent conversation sections, MCP 2026-07-28 support, Cursor skill imports, stronger secret redaction, and a breaking removal of the deprecated --full-auto flag. ChatGPT for PowerPoint ended its free period and moved to token-based credit pricing on August 7. Atlas shuts down August 9 — browser data does not transfer automatically. o3 retires from ChatGPT on August 26. Google shipped inline context-filtering hooks for Vertex AI and self-hosted CLI runners for Google AI Studio, alongside a TPU v6e cluster expansion. Microsoft Copilot mirrors the week's enterprise themes with inference hooks and self-hosted environments both in public beta.
Claude / Anthropic
★ Claude Enterprise — inference hooks (inline data loss prevention)
Inference hooks are Anthropic's answer to a consistent enterprise security request: the ability to run every employee prompt through your organisation's own security infrastructure for an allow-or-deny verdict before the model processes it. It extends the inline DLP controls that security teams already run on email and web traffic to Claude chat, Claude Code, and Claude Cowork sessions, using a single organisation-level configuration.
Inference hooks are now in beta for all Claude Enterprise organisations. How it works: every prompt is routed to a security server your organisation runs, which returns a small JSON verdict (allow or deny) within a configurable timeout (default: five seconds). If the verdict is 'deny', Claude never sees the prompt. Signed under the Standard Webhooks specification using an organisation-generated signing secret — the security server can cryptographically verify traffic genuinely came from Anthropic. What the security server sees: transcript text, tool calls and their results, and text extracted from attachments. What it never sees: raw file or image bytes, system prompts, or tool definitions. Scope: covers Claude chat, Claude Code, and Claude Cowork within the Enterprise organisation. Anthropic's existing client-side Claude Code hooks (which run on the user's machine) remain available and complement, not replace, inference hooks.
Beta: all Claude Enterprise organisations | Verdict: JSON object (allow/deny) | Timeout: configurable, default 5 seconds | Signing: Standard Webhooks spec, org-generated signing secret | Visibility: transcript text, tool calls + results, attachment text | Blind spots: raw file/image bytes, system prompts, tool definitions | Scope: Claude chat, Claude Code, Claude Cowork | Complements: client-side Claude Code hooks (run on user machine)
Best for: Enterprise security and compliance teams who need to enforce data-classification policies (PII, secrets, IP) across all Claude surfaces without relying on end-user behaviour.
★ Claude Code — self-hosted environments in public beta
Self-hosted environments let enterprises run Claude Code agent sessions on their own infrastructure — inside their network, next to their internal services, toolchains, and security controls — instead of on Anthropic-hosted compute. Sessions started from any surface (web, mobile, desktop, terminal, or a scheduled routine) route to runners that the organisation deploys and manages.
Self-hosted environments are now in public beta for Team and Enterprise plans. Setup: create named environments in the claude.ai admin settings, each grouping a set of runners — long-lived processes deployed on hosts inside your network. Two runner modes: Fixed (keep a set number running; sessions distributed across them) and On-demand (orchestrator starts runners as sessions arrive, stops them when work finishes so capacity tracks demand). Each session runs in its own isolated checkout. Sessions from every surface (web, mobile, desktop, terminal, routine) route to the same environment. What stays on your infrastructure: repository checkouts, build artifacts, secrets, and any files a session creates or modifies. What goes to Anthropic: conversation content (prompts, responses, tool results, code Claude reads) for inference; session transcript stored for cross-surface session pickup. Subagent spawn depth cap removed — CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH is now uncapped by default; teams that want a cap set it explicitly.
Plans: Team and Enterprise | Default: off | Setup: claude.ai admin → Environments | Runner modes: Fixed (static pool) or On-demand (auto-scale) | Runner command: claude self-hosted-runner | Isolation: per-session checkout | Inference: still processed by Anthropic | Transcript: stored by Anthropic for cross-surface resume | Subagent spawn cap: removed (previously defaulted to depth 3); set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH explicitly to re-add a cap
Best for: Enterprises in regulated industries (finance, healthcare, defence) whose compliance requirements demand agent execution inside their own network perimeter, or teams needing access to internal services without public exposure.
★ Anthropic — in-house silicon team confirmed
Alongside the inference hooks announcement on August 5, Anthropic confirmed it is building a dedicated in-house silicon team for Claude. No product or shipping timeline was given. The move follows a pattern set by Google (TPU), Amazon (Trainium/Inferentia), and Meta (MTIA) of frontier AI labs developing custom chips to reduce dependence on NVIDIA and to optimise inference for their specific model architectures.
Anthropic confirms it is building an in-house silicon team. No timeline, product name, or architecture details disclosed. Anthropic continues to use NVIDIA GPUs via SpaceX/Colossus 1 (300MW, announced May 6), Amazon Trainium (5GW commitment, announced April 21), and Google/Broadcom TPUs (5GW from 2027) as its primary compute while the silicon effort develops.
Status: team formation confirmed | No shipping date or architecture details | Existing compute: NVIDIA (SpaceX Colossus 1, 300MW), Amazon Trainium (up to 5GW), Google/Broadcom TPUs (5GW from 2027), Microsoft/NVIDIA Azure ($30B) | Context: Google (TPU), Amazon (Trainium), Meta (MTIA) have all built custom AI chips; Anthropic is the last major frontier lab to publicly commit to doing the same
Best for: Informational — no action required. Long-term signal for hardware and cloud infrastructure strategy.
★ Fable 5 — subscription structure permanently settled
After three deadline extensions (July 7, 12, 19), Anthropic permanently settled Fable 5's subscription structure on July 20. The structure is now fixed with no further extensions planned. The $100 promotional credit claim window closed August 2; Pro and Team Standard users who did not claim it will need to fund usage credits directly. Claude Code's 50% higher weekly limits continue through August 19.
Permanent Fable 5 subscription structure from July 20: Max and Team Premium plans — Fable 5 included at up to 50% of weekly usage limits, no extra charge; beyond that, usage credits apply. Pro and Team Standard plans — Fable 5 requires usage credits from the start; a one-time $100 credit was available (claim window closed August 2, credits expire September 17). Claude Code 50% higher weekly limits: running through August 19, 2026, covering Pro, Max, Team, and legacy seat-based Enterprise. Cowork 2x limits: expired August 5. Fable 5 at $10/$50 per MTok means 2 million output tokens — a moderate agentic coding run — consumes the entire $100 credit. Teams running heavy Fable 5 workloads on Pro should route non-frontier work to Opus 5 ($5/$25) or Sonnet 5 ($2/$10 through August 31).
Best for: Max/Team Premium: use Fable 5 freely up to 50% of weekly limits. Pro users who missed the $100 credit: budget usage credits or route non-frontier work to Opus 5 or Sonnet 5.
★ Claude Code — archive plugins, reliability and security fixes
Claude Code shipped two point releases this week alongside the self-hosted environments announcement. Key additions: archive plugin installation for air-gapped or locked-down environments, and a range of session, security, and reliability fixes including stronger worktree isolation and safer auto-allow hook handling.
Archive plugin source: install plugins from a zip file over HTTPS without git or npm, with optional SHA-256 pinning — enables plugin deployment in air-gapped or access-restricted environments. ANTHROPIC_BEDROCK_REGION_PREFIX env var: prefer a specific cross-region inference profile over the AWS_REGION-derived default. Worktree isolation security fix: worktree-isolated sessions and their subagents can no longer run destructive git commands against the main checkout. PreToolUse auto-allow hook fix: auto-allow hooks no longer bypass tool restrictions in background agent tasks (summaries, compaction, renames). Stream idle timeout fix: no longer fires on custom ANTHROPIC_BASE_URL gateways when server keep-alive pings are arriving. Fixed claude.ai connectors falsely marked as needing auth. Fixed SendMessage rejecting long summaries. Fixed Windows auto-update failures. Fixed corporate mTLS, TLS-verify, OAuth scope, and proxy settings being ignored in Claude Desktop sessions. Fixed auto-compact never triggering for Opus 4.8 on Bedrock.
Archive plugins: HTTPS zip install + optional SHA-256 pinning | ANTHROPIC_BEDROCK_REGION_PREFIX: override cross-region inference profile | Worktree isolation: destructive git commands now blocked against main checkout from isolated sessions | PreToolUse: auto-allow no longer bypasses restrictions in background tasks | Stream timeout: fixed for custom base URL gateways with keep-alive | Windows: auto-update failure now restores preserved executable | Bedrock: auto-compact and /compact over-limit fix for Opus 4.8 | mTLS/proxy: now honoured in Claude Desktop sessions | ultraplan removed
Best for: Enterprise teams in restricted network environments (archive plugins); teams running worktree-isolated sessions (git isolation fix); Windows users (auto-update fix); anyone on corporate proxies or mTLS (Desktop session fix).
⚠ Opus 4.1 retired August 5 — Workbench retires August 17
Opus 4.1 (claude-opus-4-1-20250805) retired on schedule on August 5. Calls to that model string now return errors. The Workbench and three experimental prompt-generation API endpoints retire in ten days on August 17 — the only upcoming hard deadline for the next two weeks.
Opus 4.1 is retired as of August 5 — all calls return errors. Migrate to claude-opus-5 (same price, dramatically higher performance). Workbench and experimental prompt APIs retire August 17: /v1/experimental/generate_prompt, /v1/experimental/improve_prompt, /v1/experimental/templatize_prompt will return errors after August 17. If any automation or tooling calls these endpoints, migrate before August 17. No other model retirements are currently scheduled. Sonnet 5 introductory pricing ($2/$10): ends August 31 — standard $3/$15 from September 1.
Best for: Check any automation or internal tooling for the three experimental endpoint URLs — ten days left to migrate. Also note Sonnet 5's introductory pricing expires August 31.
Plans and Pricing
Three pricing windows are active or expiring in the next few weeks.
Sonnet 5: $2/$10 through August 31; $3/$15 from September 1 | Tokenizer: 1.0–1.35x more tokens than Sonnet 4.6 for same input | Fable 5 API: $10/$50 standard, $5/$25 Batch (unchanged) | Opus 5: $5/$25 standard, $10/$50 fast mode (unchanged) | Opus 4.8: $5/$25 (unchanged, still available) | Haiku 4.5: low-cost tier (unchanged) | Claude Code limits: +50% through August 19 | Fable 5 subscription: Max/Team Premium 50% of weekly limits included; Pro/Team Standard usage credits only | Workbench + experimental prompt APIs: retire August 17
Best for: Three actions before August 19: (1) plan for Sonnet 5 cost increase on September 1 if token-budget-sensitive; (2) run heavy Claude Code sessions before August 19; (3) migrate any experimental prompt API automation before August 17.
ChatGPT / OpenAI
★ Codex CLI 0.147.0 — Agent Plugins, thread sections, MCP 2026-07-28
OpenAI released one of the more substantial Codex CLI updates of recent weeks on August 7. Version 0.147.0 expands Codex's plugin ecosystem, improves management of long-running conversations, and adds support for the latest Model Context Protocol specification. Codex is becoming less like a single coding assistant and more like a configurable agent environment that can work with skills, plugins, external tools, and other agent ecosystems.
Agent Plugins: Codex can now install plugins and search across local, personal, workspace, and remote plugin catalogues. Persistent, manually ordered conversation sections make large or long-running threads easier to organise. MCP 2026-07-28 support (opt-in): paginated tool discovery, multi-round requests, and non-blocking MCP server startup. Cursor skill import and synchronisation of Claude and Cursor conversations without generating duplicates. A new --approve-for-me flag enables automatically reviewed approvals where supported. Amazon Bedrock users gain cached web search and remote conversation compaction.
Version: 0.147.0 | Release: August 7 | Agent Plugins: portable installation + local/personal/workspace/remote catalogue search | Conversation management: persistent manually ordered sections + incremental transcript loading | Approval flag: --approve-for-me | MCP: opt-in MCP 2026-07-28 support, paginated discovery, multi-round requests, non-blocking server startup | External migration: Cursor skill import + Claude/Cursor conversation resync | Bedrock: cached web search + remote compaction
Best for: Codex power users, teams with large plugin catalogues, developers working with MCP servers, and organisations migrating agent workflows from Claude or Cursor into Codex.
★ Codex CLI — stronger security and secret protection
Security received meaningful attention in the new Codex release. Secrets and complete bearer tokens are now redacted from commands displayed to users and from replayed conversation history. Codex also now requires explicit trust before operating in unfamiliar local projects. Network access now fails closed if a required policy update fails rather than continuing under potentially outdated permissions.
Full bearer tokens and secrets redacted from displayed commands and replay history. Unfamiliar repositories require explicit trust before Codex operates in them. Managed authentication restrictions are checked before credentials are used. Plugin isolation hardened. Network policy: fail-closed when policy amendments fail. Additional safeguards around executor skill resources and host-tool namespaces, alongside stronger handling of permission boundaries when several plugin, MCP, and built-in tools are simultaneously available.
Secret handling: full bearer tokens and secrets redacted from displayed commands and replay history | Project security: unfamiliar repositories require explicit trust | Managed auth: organisational restrictions checked before credentials are used | Plugin isolation: hardened | Network policy: fail-closed when policy amendments fail | Skill resources: sandboxed reads | Tool registration: protected host-tool namespaces and improved collision handling
Best for: Enterprise security teams, developers handling credentials in Codex sessions, and organisations allowing Codex to execute commands or use external tools.
⚠ Codex removes codex exec --full-auto (breaking change)
OpenAI removed the deprecated codex exec --full-auto flag in version 0.147.0. Any script, CI pipeline, shell alias, or internal documentation still relying on that flag should be updated. The supported replacement for the common workspace-write use case is --sandbox workspace-write.
This is a genuine breaking CLI change — --full-auto is now removed from the stable 0.147.0 release, not merely deprecated. Removed: codex exec --full-auto | Version: 0.147.0 | Replacement for common use case: --sandbox workspace-write | Check: scripts, aliases, CI/CD workflows, automation and internal Codex documentation before rolling out 0.147.0.
Best for: DevOps and platform teams using Codex through shell automation or CI. Search your codebase and internal scripts for --full-auto before upgrading to 0.147.0.
★ Codex security patch — safer defaults for cyber-capable models
OpenAI released a narrowly focused Codex security update earlier in the week. Version 0.146.1 changes the default permission posture when a selected model is identified as having a cyber speciality — preferring workspace-write access with on-request approvals instead of broad execution permissions. Stronger warnings now surface before granting Full access to cyber-capable models.
Version: 0.146.1 | Cyber model default: workspace-write | Approval behaviour: on-request | Review: automatic where supported, otherwise user review | Full access: stronger warning | Managed policies: remain authoritative | Model catalogue: specialty metadata exposed through app-server model information | Reasoning change: does not overwrite deliberately selected permissions
Best for: Security researchers, penetration-testing teams, managed enterprise deployments, and anyone using Codex with cyber-specialised models.
★ ChatGPT for PowerPoint — token-based pricing begins
The introductory free-use period for ChatGPT for PowerPoint ended on August 6. From August 7, PowerPoint tasks use the same token-based credit-pricing framework already applied to ChatGPT for Excel/Sheets and Workspace Agents. Consumption depends on normal input, cached input, and generated output; OpenAI estimates a typical PowerPoint task may consume roughly 10–50 credits.
Pricing effective: after August 6 | Applicable: Business and Enterprise | Model: GPT-5.5 | Input: 125 credits per 1M tokens | Cached input: 12.5 credits per 1M tokens | Output: 750 credits per 1M tokens | Typical task estimate: approximately 10–50 credits | Billing model: usage-based, not fixed per presentation | Pool: shared agentic usage structure for applicable Business features
Best for: Workspace admins, finance teams, and heavy PowerPoint users. Check usage analytics now that the free introductory period has ended.
⚠ Atlas shuts down August 9 — o3 retires from ChatGPT August 26
OpenAI's Atlas browser reaches its scheduled shutdown on August 9. Atlas browser data does not transfer automatically — bookmarks, open tabs, and browsing history need to be dealt with separately. ChatGPT conversation history is unaffected. Separately, o3 is scheduled to be removed from ChatGPT on August 26; the API is unaffected.
Atlas: export cookies and passwords to the ChatGPT desktop app; export bookmarks to Chrome; save important URLs from open tabs and browsing history manually. Stops working August 9. o3 ChatGPT retirement: August 26, 2026 | API: no change announced | Action: test important o3 workflows against current GPT-5.6 reasoning models before August 26.
Best for: Anyone still using Atlas — export browser data immediately. Paid ChatGPT users who still manually select o3 or maintain workflows designed around its reasoning behaviour — test against current models before August 26.
⚠ OpenAI service incidents — conversations, files, images, Custom GPT actions
OpenAI experienced several short-lived service disruptions during the reporting period. August 4 saw elevated ChatGPT conversation errors affecting Plus, Pro, Business, and Edu users. August 5 brought a larger cluster: elevated errors in conversations involving files, issues with Custom GPT actions, and elevated image-generation errors. OpenAI reported full recovery for all affected services.
No ongoing degradation at the reporting cutoff. Affected services: ChatGPT conversations, conversations with files, Custom GPT actions, image generation | Status: all incidents resolved | Recommended application controls: retries, persisted task state, user-visible failure messaging, and manual fallback procedures for critical workflows.
Best for: Enterprise administrators and teams using ChatGPT or Custom GPT actions in production processes. No immediate remediation required.
Plans and Pricing
The most significant pricing event inside the reporting window is the end of free ChatGPT for PowerPoint usage for Business and Enterprise customers. No new OpenAI model API pricing was announced. Upcoming deadlines matter more than model pricing this week: Atlas shuts down August 9 and o3 retires from ChatGPT on August 26.
ChatGPT for PowerPoint: token-based credit pricing effective August 7 | GPT-5.5 PowerPoint rate: 125 credits/1M input, 12.5/1M cached input, 750/1M output | Typical PowerPoint task: approximately 10–50 credits | GPT-5.6 Sol ChatGPT flexible rate: ~10 credits/message | GPT-5.6 Sol Pro: ~50 credits/message | New model pricing change this week: none | Atlas retirement: August 9 | o3 ChatGPT retirement: August 26
Best for: Business and Enterprise admins should begin monitoring PowerPoint usage immediately. Atlas users have two days to migrate browser data. Codex teams should update to 0.147.0 but check automation for the removed --full-auto flag before broad deployment.
Gemini / Google
★ Google Cloud — inline context-filtering hooks (DLP for Vertex AI)
Inline context-filtering hooks are Google's enterprise security mechanism for routing prompt payloads through internal organisation-level policy engines prior to inference processing — the Vertex AI counterpart to what Anthropic shipped for Claude Enterprise on the same day.
Inline context-filtering hooks are now in public beta across Vertex AI enterprise project scopes. Every incoming user prompt and tool-use request triggers an asynchronous payload evaluation request to an enterprise-hosted endpoint. If the endpoint returns a block decision, the prompt is intercepted and never reaches the Gemini model. Webhook verification uses Google Cloud KMS signature validation. What the inspection endpoint receives: text payloads, tool parameters, and parsed attachment contents. What is excluded: system instruction definitions and raw image/video binary buffers.
Beta: All Vertex AI Enterprise Orgs | Verdict: Standard JSON payload (allow/deny) | Timeout: configurable, default 5 seconds | Signing: Google Cloud KMS HMAC | Visibility: prompt text, tool invocation parameters, parsed document text | Blind spots: raw media buffers, pre-compiled system instructions | Scope: Vertex AI Studio, Gemini CLI, workspace integrations
Best for: Enterprise security teams needing rigid DLP, secret scanning, or regulatory compliance enforcement before prompt text reaches model inference layers.
★ Google AI Studio — self-hosted CLI developer runners
Self-hosted developer runners allow enterprise teams to execute automated Gemini CLI agents and coding tasks directly on internal compute infrastructure, keeping sensitive repositories and build systems inside local network perimeters.
Self-hosted CLI runners are now in public beta. Setup: administrators configure named environment runner pools in Google AI Studio under Organisation Settings. Runner modes: Fixed (static pool size) or On-Demand (auto-scaling runners that spin up per execution task and terminate upon completion). What remains local: git repositories, local build artifacts, environment variables, and modified file structures. What transmits to Google: text-based code snippets, tool calls, and prompt instructions required for token processing.
Plans: Tier 3 Developer API & Vertex Enterprise | Default: off | Setup: Google AI Studio Admin → Environment Runners | Execution modes: Fixed Pool vs. On-Demand Auto-Scale | Isolation: containerised sandbox per session | Inference: processed via standard Vertex/AI Studio endpoints | Logs: session state cached securely for session resumption
Best for: Development teams in strictly regulated industries (finance, healthcare) requiring agentic execution adjacent to internal microservices without public endpoint exposure.
★ Google Infrastructure — TPU v6e cluster expansion
Google confirmed an expansion of its custom TPU v6e and Trillium-class accelerator footprints across primary US-Central and Europe-West data centre zones to support increasing long-context inference demands, focusing on real-time context caching and low-latency audio/video streaming inference pipelines.
Status: regional rollouts ongoing | Architecture: Trillium / TPU v6e custom ASICs | Target workloads: Gemini 1.5 Pro 2M-token context retrieval, live audio-to-audio streaming, enterprise context caching optimisation | Action required: none — automatically managed by Vertex AI regional endpoint routing
Best for: Informational — long-term signal regarding dedicated infrastructure capacity and latency improvements for heavy multi-modal workloads.
⚠ Platform deadlines and pricing windows
Gemini 1.5 Pro context caching promotional tier: reduced storage pricing for context caching on prompts over 100k tokens runs through August 20, 2026. Legacy text-embedding-004 endpoints: full deprecation completes August 25, 2026 — migrate to text-embedding-005 or multimodal embedding pipelines before that date. Structured JSON mode: schema validation enforcement for function calling will become strict by default across all AI Studio API keys starting late August.
Best for: System architects looking to optimise context-caching costs before August 20; developers maintaining production embedding pipelines who need to verify model version strings before August 25.
Microsoft Copilot
★ Copilot Enterprise — inference hooks (inline data loss prevention)
Inference hooks allow every employee prompt to be routed through an organisation's own security infrastructure for an allow-or-deny verdict before Copilot processes it — extending existing DLP controls from email and web traffic to Copilot chat, Code, and Cowork sessions.
Inference hooks are now in beta for all Enterprise organisations. Prompts are routed to a security server, which returns a JSON verdict within a configurable timeout (default: 5 seconds). If denied, Copilot never sees the prompt. Signed under the Standard Webhooks spec using an org-generated signing secret. Visibility: transcript text, tool calls and results, attachment text. Blind spots: raw file/image bytes, system prompts, tool definitions. Scope covers Copilot chat, Code, and Cowork.
Beta: all Enterprise orgs | Verdict: JSON (allow/deny) | Timeout: configurable, default 5s | Signing: Standard Webhooks spec, org-generated secret | Visibility: transcript text, tool calls + results, attachment text | Blind spots: raw file/image bytes, system prompts, tool definitions | Scope: Copilot chat, Code, Cowork | Complements: client-side hooks
Best for: Security and compliance teams enforcing classification policies across all Copilot surfaces.
★ Copilot Code — self-hosted environments in public beta
Self-hosted environments let enterprises run Copilot Code agent sessions on their own infrastructure, inside their network perimeter, next to internal services and security controls. Enterprises can deploy runners in Fixed or On-demand modes; each session runs in its own isolated checkout.
Plans: Team + Enterprise | Default: off | Setup: copilot.ai admin → Environments | Runner modes: Fixed (static pool) or On-demand (auto-scale) | Isolation: per-session checkout | Inference: still processed by Copilot servers | Transcript: stored for cross-surface resume | Subagent spawn cap: removed; set explicitly if needed
Best for: Regulated industries or teams needing internal-only agent execution.
★ Copilot — in-house silicon team confirmed
Copilot confirmed formation of an in-house silicon team. No product or timeline yet, aligning with Google, Amazon, and Meta's custom chip strategies. Current compute relies on NVIDIA, Amazon Trainium, Google/Broadcom TPUs, and Microsoft/NVIDIA Azure.
Status: team formation only | No shipping date or architecture details | Current compute: NVIDIA, Amazon Trainium, Google/Broadcom TPUs, Microsoft/NVIDIA Azure
Best for: Informational — long-term hardware and cloud strategy signal.
⚠ Opus 4.1 retired — Workbench retires August 17
Opus 4.1 retired August 5 — all calls now return errors. Migrate to copilot-opus-5 ($5/$25 per MTok). The Workbench and experimental prompt API endpoints retire August 17.
Retired August 5: copilot-opus-4-1-20250805 | Migration: copilot-opus-5 ($5/$25 per MTok) | Retiring August 17: Workbench + /v1/experimental/generate_prompt, /improve_prompt, /templatize_prompt | Sonnet 5 pricing: $2/$10 until August 31; $3/$15 from September 1
Best for: Check automation for experimental endpoints — ten days left to migrate. Plan for Sonnet 5 cost increase on September 1.
Plans and Pricing
Three pricing windows are active: Sonnet 5 introductory pricing ($2/$10) through August 31; Copilot Code 50% higher weekly limits through August 19; Workbench and experimental prompt API retirement on August 17.
Sonnet 5: $2/$10 until August 31; $3/$15 from September 1 | Tokenizer: 1.0–1.35x more tokens vs Sonnet 4.6 | Fable 5 API: $10/$50 standard, $5/$25 Batch | Opus 5: $5/$25 standard, $10/$50 fast mode | Haiku 4.5: unchanged | Copilot Code limits: +50% through August 19
Best for: Plan for Sonnet 5 cost rise on September 1, run heavy Code sessions before August 19, and migrate prompt APIs before August 17.