That was the notice window Anthropic gave developers before retiring the legacy Workbench and three experimental prompt-engineering endpoints. August 17 was the date. The endpoints return errors now. If you had saved prompts, variables, or eval configurations in the old Workbench and missed the export window, Anthropic says there is no recovery path. Not limited support. Not a grace-period ticket. Gone.
The replacement is stateless and browser-local. It is also, deliberately, lighter. Whether that trade lands well depends entirely on how much you had built on what got cut.
Elsewhere on the frontier: OpenAI shipped a teen-safety mode for ChatGPT on the same day. Age prediction routes users estimated to be 13 to 17 into a filtered experience with study tools and break reminders. Four years after ChatGPT launched and teenagers immediately became its heaviest consumer cohort.
On August 17, Anthropic closed the legacy Claude API Workbench and retired the three endpoints that powered its prompt-engineering workflows: /v1/experimental/generate_prompt, /v1/experimental/improve_prompt, and /v1/experimental/templatize_prompt. Requests to those paths return errors now. The announcement came July 17, giving developers exactly 31 days to export saved data and migrate any pipelines calling the endpoints.
Anthropic's documented deprecation policy for models requires a minimum of 60 days notice. The Workbench is not a model. The endpoints were labeled experimental, which carries no SLA and no long-term commitment. The shorter window was within Anthropic's stated rights. Whether it matched the reasonable expectations of teams who had built production pipelines on those endpoints is a different question entirely.
The blast radius lands in two places. First: any pipeline that was automating prompt generation, improvement, or templatization through the API is broken as of yesterday. Migration means moving that logic to direct Messages API calls, likely with a system prompt that approximates what the experimental endpoint was doing. Second: any team using the Workbench as a shared prompt repository either exported before the deadline or lost the data. No recovery path. Not limited support. Gone.
The replacement Workbench is simpler: stateless, browser-local, no server-side storage. Test prompts, export them as code, close the tab. This is a cleaner architectural decision for a company managing data at Anthropic's scale, and it removes a class of liability around storing and securing developer prompt assets. It is also, deliberately, not a prompt registry. That affordance was cut along with the endpoints.
The pattern sits inside a larger platform direction. This quarter, Anthropic shipped support for the MCP 2026-07-28 specification, making Claude the first major model client to adopt the stateless-core, hardened OAuth 2.0 and OIDC version of the protocol, with versioned extensions for Apps and Tasks and new connector features including embedded UI, enterprise-managed auth, observability, and private network tunnels. It also deprecated older streaming formats and is now cutting the legacy prompt-tooling surface. The direction is coherent: stateless, explicit-control, agentic-first. The 2022-era managed-prompt-loop affordance does not fit that roadmap, and it was removed on a schedule that prioritized platform velocity over developer migration comfort.
The read: Anthropic accepted meaningful developer friction to move the platform faster. That choice is defensible on its merits. What it is not is invisible. The 31-day window on a permanent deletion, on a week with no major product launch to offset it, is a signal about where developer workflow stability sits in the priority order when platform modernization is moving.
Anthropic pushed v2.1.234 on August 17 and v2.1.235 on August 18. The cadence is routine; the contents are not thin.
V2.1.234 adds GitLab MR badge support. Repos with a GitLab remote and an authenticated glab CLI now show the merge request number in the footer and statusline, with draft, pending, and green states visible at a glance. GitHub parity has existed for a while; GitLab-heavy engineering organizations get the same in-editor context now. The other notable addition is automatic session continuation when usage limits reset: Claude Code resumes the session when the usage window rolls over rather than requiring a manual restart. For anyone running long agentic tasks on capped plans, this removes a silent failure mode that previously meant a task either completed incorrectly or had to be started over entirely. Remote Control file upload and the /resume session-history leak are both correctness fixes.
V2.1.235, today, adds an optional spellcheck setting that underlines misspelled words in the prompt input and fixes a whole-prompt-cache invalidation bug triggered when language servers disconnected. The cache bug has direct cost implications at scale: prompts that should have been served from cache were being re-processed. Fix it, reclaim efficiency.
Builder's move: claude update. The cache fix and session continuation are both substantive improvements for production agentic workflows.
OpenAI launched ChatGPT for Teens on August 18. Users estimated to be between 13 and 17, detected via account signals and behavioral prediction rather than hard verification, are routed automatically into a filtered mode. The mode blocks suicide and self-harm topics, romantic and sexual content, and certain high-risk interaction patterns. Study Mode guides through problem-solving step by step rather than providing direct answers. Regular break reminders surface in-session. Parental safety notifications can be triggered for high-risk interactions, reviewed by human personnel before sending.
The TechCrunch headline reads: "OpenAI launches a safer ChatGPT for teens, years after teens started using it." That is accurate. ChatGPT launched November 2022. Teen mode launched August 2026.
The contrast against Anthropic is worth naming: Anthropic has maintained conservative platform defaults and account-gated access since launch and has never run a fully open consumer product without minimum age terms. The access models differ, which makes a direct comparison imprecise. What is comparable is the sequencing. One lab built consumer scale first and is retrofitting safety infrastructure. The EU AI Act's fine-enforcement provisions activated August 2, 2026. The gap between those two dates and this announcement is probably not a coincidence.
The execution caveat: age prediction without hard verification. OpenAI routes users it estimates are minors into teen mode. At scale, without a verified identity system, accounts that misrepresent age at signup are not covered by that routing. The guardrail is probabilistic, not categorical. That gap between "we have teen mode" and "teens are in teen mode" is where the actual safety question lives.
On August 17, xAI pushed Grok Image 2.0 into the Grok mobile app. No API release, no pricing disclosure, no technical paper. An app update that adds a new image generation model to the Grok consumer experience.
On the same day, Google deprecated its legacy standalone image-generation endpoints in the Gemini API. The deprecation is a cleanup, not a retreat: image capability has migrated to Gemini 3.7 Flash's native multimodal output, which became generally available August 13. Google is consolidating surfaces.
The juxtaposition: xAI is adding a new consumer image surface. Google is retiring an old API in favor of a better-integrated model. Both moves point at the same conclusion: image generation is no longer a standalone feature, it is a modality inside a general-purpose model. xAI is building toward that integration. Google is already there and removing the scaffolding.
The practical note for builders: Gemini 3.7 Flash's native image output is GA, priced ($0.75 per million input tokens, $3.75 per million output tokens through December), and available through the standard Gemini API. Grok Image 2.0 is in the mobile app with no disclosed API path. If you need image generation through an API today, Google's is the only new-generation option with a clear pricing and availability signal.
Grok 4.6 promotional window closes August 19. Doubled usage credits in Grok Build and Cursor end tonight for most accounts. Grok 4.6, released August 12 and available in GitHub Copilot across eight development surfaces since August 14, is built for long-running agents and complex multi-step workflows. The promotional window was xAI's evaluation play. That play ends tonight.
GPT-5.6 Sol Ultrafast (Cerebras) remains in limited API preview. The 750 tokens-per-second tier, powered by Cerebras wafer-scale chips, has been in selective preview since August 13. Fourteen times standard throughput, no measurable quality drop on benchmark evaluations. No pricing disclosed, no date for broader access.
Gemini 3.7 Flash is the established workhorse. GA since August 13, 1M context window, strong gains on software engineering benchmarks, introductory pricing through December. Gemini Robotics ER 2, for advanced spatial reasoning and multi-robot coordination, is in public preview.
Mistral: no August-specific release, but the industrial stack is moving. OCR 4.1 (July 16) is current; the Airbus, BMW, and ASML enterprise deployments are in production. The European sovereign-AI positioning is the ongoing story, sharpened by EU AI Act fine enforcement that activated August 2.
claude update or reinstall. Enable spellcheck in settings to get inline spell checking in the prompt input. The cache fix is automatic and applies to all sessions with language server integrations active.claude update. Set CLAUDE_CODE_PROJECT_DIR_NAME to customize the project name shown in the statusline. GitLab MR badges require an authenticated glab CLI with repo access.Daily digest at 9 PM ET. Weekly magazine every Friday morning. Six labs, one feed. No spam, one-click unsubscribe.