Haiku 3 Retires Today — Last Call for claude-3-haiku-20240307
Today is the hard cutoff. After April 19, API calls against claude-3-haiku-20240307 fail with a 404-class error, not a soft fallback or deprecation warning. Anthropic has been telegraphing this for six weeks, but plenty of production systems — RAG pipelines, ETL jobs, agent loops, legacy Zapier flows, random weekend scripts nobody remembers writing — are still pinned to that string. If you run ops anywhere that touched Claude in 2024, check one more time before Monday morning.
Migration target is Claude Haiku 4.5 (claude-haiku-4-5-20251001) — same price tier, materially faster, better eval scores across the board, and it supports vision and tool use without caveats. The next dates on the retirement calendar: the 1M-token context beta for Sonnet 4.5 and Sonnet 4 expires April 30, and Opus 4 plus Sonnet 4 both fully retire June 15. Migration targets are Opus 4.7 and Sonnet 4.6.
Claude Cowork Goes Generally Available — Analytics, OpenTelemetry, and RBAC
Claude Cowork — the desktop agent that opens files, runs dev tools, clicks through apps, and drives a computer on the user’s behalf — is now GA on macOS and Windows inside the Claude Desktop app, out of research preview. The Enterprise bundle that ships with GA is the interesting part: usage analytics via the Analytics API, OpenTelemetry support for observability pipelines, and role-based access controls that admins can wire up manually or via SCIM from their identity provider. Groups get custom roles defining which Claude capabilities they can use.
Two more Cowork pieces rolled alongside. A new Customize panel inside Claude Desktop groups skills, plugins, and connectors in one place so users don’t have to hunt. And Pro/Max users get a persistent agent thread from Claude Desktop or the iOS/Android apps to manage Cowork tasks on the go — kick off a task from your phone, let it run on your laptop. Scheduled recurring tasks are now first-class. Translation: Cowork isn’t a tech demo anymore. It’s a second shipping surface next to Claude Code.
Claude Design Weekend Momentum — Export-to-Canva Pipeline Draws First Real Reviews
Claude Design, which shipped Friday, had its first weekend in the wild. Early reviews from InsideHook, gHacks, and IBL News call out the same pattern: the prompt-to-prototype loop is genuinely good for pitch decks, wireframes, and marketing one-pagers, and the Canva export path means teams don’t have to leave a tool they already trust. The novel capability nobody was talking about at launch is that Claude Design can ingest an existing codebase or Figma library and apply the team’s actual design system to every asset it generates — the thing that always took a designer half a day.
The market reaction that started Friday (Figma -7%, Adobe -4%) stuck through the weekend. InsideHook’s framing was the most direct: “the internet exploded” over Claude Design because Anthropic showed up in the one category where AI was supposed to be augmenting designers, not replacing the tool.
Plugin Marketplace Crosses 4,200 Skills and 2,500 Marketplaces
The Claude Code plugin ecosystem is now at 4,200+ skills and 2,500+ marketplaces, per data roundups this week from Felo, BuildToLaunch, and ClaudeMarketplaces. The official claude-plugins-official marketplace now carries 101 plugins — 33 built by Anthropic (language servers, dev workflow, setup utilities) and 68 partner plugins from GitHub, Playwright, Supabase, Figma, Vercel, Linear, Sentry, and Stripe. Team and Enterprise admins got new controls earlier this month to gate which plugins an org can install.
Thoughtworks added the Claude Code plugin marketplace to its Technology Radar in the latest update — a reasonable proxy for when enterprise architects start taking something seriously. If you’re evaluating platform-level LLM tooling for a large team, the decision surface is no longer just the model; it’s which ecosystem’s plugins and marketplaces match the way your team already works.
Xcode 26.3 Ships With Native Claude Agent SDK Integration
Apple and Anthropic pushed a joint announcement this week that Xcode 26.3 now ships with the Claude Agent SDK integrated natively — the same underlying harness that powers Claude Code, but surfaced directly in the IDE. Developers get subagents, background tasks, plugins, and routines without leaving Xcode. For any team that went hybrid Swift/Kotlin or ships iOS products from a Mac-centric shop, this collapses what was a two-window workflow (IDE plus Claude Code terminal) into one.
This is also the second IDE integration Anthropic has shipped this quarter — following the Zed, JetBrains, and VS Code work earlier — and the first where a platform vendor chose to bake Claude in by default instead of publishing it as an optional extension. Apple is implicitly picking a horse in the coding-agent race.
Auditing for Haiku 3 Today — the 5-Minute Version
If you’re scrambling this morning, a quick grep is usually enough. Search your codebase and config stores for claude-3-haiku-20240307 and claude-3-haiku. Check environment variables in production, Vercel/Netlify deploy configs, Zapier and n8n workflows, Airflow DAGs, Lambda env vars, Kubernetes ConfigMaps, and any docs or READMEs your team copy-pastes from. Then run a test call against each updated string before the end of the day.
A lot of the Haiku 3 dependencies out in the wild aren’t in your repo — they’re in third-party tools that someone on your team wired up two years ago and forgot about. If you maintain any Claude integrations for customers, a proactive “we’ve already migrated you” note sent today lands a lot better than a Tuesday fire drill.
Mainstream Press Piles On Over the “Effort Nerf” — This One Isn’t Going Away
What started as Reddit and Hacker News complaints three weeks ago is now a mainstream narrative. Fortune ran a long piece on user backlash over reduced Claude performance. Axios framed it bluntly: “Anthropic’s AI downgrade stings power users.” The Register’s headline — “Claude is getting worse, according to Claude” — went around every devtools channel on Slack. The core issue: Anthropic quietly lowered Claude’s default effort level to save tokens, didn’t announce it widely, and Cursor/Claude Code users noticed the regression before Anthropic confirmed it.
Opus 4.7’s new xhigh effort level, which shipped Thursday, reads differently in this light — less a new capability, more the cap on a throttle that shouldn’t have been quietly installed in the first place. Enterprise contacts are now asking pricing questions on calls: if default effort can drop without notice, what does the seat I’m paying for actually entitle me to? The Opus 4.7 task-budget feature is partly the technical answer. The communication answer still isn’t there.
Accenture, PwC, Snowflake — The Enterprise Delivery Layer Solidifies
Three partnership announcements are worth watching as a set, not individually. Accenture and Anthropic expanded their multi-year deal to move enterprises from AI pilots to full deployment (Accenture is now the largest internal-Claude shop in the world). PwC US and Anthropic announced a collaboration to push “AI-Native” enterprise plugins into finance and healthcare & life sciences — two of the most heavily-regulated categories, where model-only vendors historically stalled. Snowflake and Anthropic expanded their partnership with a $200M commitment for agentic AI on Snowflake’s data cloud.
The pattern: Anthropic isn’t building a services arm, but it’s wiring up the three biggest delivery channels (Big 4 consulting, global systems integrator, enterprise data platform) so customers have a paved path from procurement to production. Combined with last month’s $100M Claude Partner Network announcement — 5× partner team — this is the enterprise motion that usually precedes a serious IPO roadshow.
Anthropic Is the Software-Sector Disruption Story of Q2
Two pieces from the finance press framed the same point this week. The Motley Fool ran a long analysis on “how Anthropic has wiped out trillions from the software sector” — the cumulative drag on SaaS multiples as investors reprice application-layer companies against a world where AI agents can assemble equivalent functionality on demand. MarketWise’s IPO preview went deeper on the thesis: if Claude Design can produce a serviceable Figma alternative on day one, the question isn’t whether AI disrupts SaaS — it’s which categories survive and at what margin.
Figma is the obvious name. But the piece’s deeper list — project management, note-taking, low-code/no-code, BI dashboards, customer support ticketing — is the category map for every public SaaS CFO presenting into Q2 earnings next month. The multiples compression isn’t hypothetical anymore.
The “Effort Nerf” Is Anthropic’s First Real Governance Test
Every hyperscaler eventually learns the same lesson: at a certain revenue scale, unilateral changes to defaults stop being engineering decisions and become trust decisions. Anthropic quietly dropped Claude’s default effort level to save tokens. It was probably the right call on compute margins. It was unquestionably the wrong call on disclosure. Heavy users noticed within days — Cursor traces, Claude Code sessions, agent benchmarks all showed the regression before Anthropic confirmed it. The mainstream coverage came three weeks later, which means the enterprise procurement team reading Fortune this morning is asking questions that Anthropic sales reps haven’t been briefed on.
The technical fix shipped Thursday with Opus 4.7 — explicit xhigh effort control, per-task budgets, effort telemetry in the API. The organizational fix hasn’t shipped. What heavy users want isn’t just a new dial — it’s a commitment that default behavior won’t change again without a changelog entry. If you’re selling agents that run autonomously for hours, you cannot quietly change how hard they think. The next five days will show whether Anthropic treats this as a product bug or a policy one. The IPO banker modeling $800B is watching the same thing.