Code with Claude SF kicked off this morning at the Letterman Digital Arts Center, and if you spent yesterday reading the predictions, you’d have braced for five separate launches landing on stage. The honest read, eight hours in: most of those launches actually shipped weeks ago. The one genuinely new thing is buried in a Claude Code CLI bump that posted overnight.
This is a same-day debrief written for engineering managers and IT buyers — the people who’ll spend Thursday morning fielding questions like “did Sonnet 4.8 launch?” and “is auto mode safe to turn on now?” Short answer: not yet, and yes — but it’s been safe to turn on since March.
What actually landed today (May 6, 2026)
Three things, ordered by what actually moves a roadmap.
1. Claude Code CLI 2.1.129. Posted around midnight Pacific. The version bump itself is small, but the gateway model discovery additions are the most interesting line item Anthropic has shipped this week. The release adds a --plugin-url flag that pulls a plugin .zip from any URL for session-only installs (no global registry write — useful for one-off tooling tests), exposes a new env var CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY that lets eligible orgs surface preview models in their CLI model list, and registers a claude-mythos-preview model identifier behind that gateway. There are also fifteen-plus fixes covering OAuth refresh, MCP tool resolution, and cache invalidation.
The Mythos line item matters for one specific audience: orgs already inside Project Glasswing’s defensive cybersecurity preview (the gated April 7 release with AWS, Apple, Microsoft, Google, NVIDIA, CrowdStrike, and the Linux Foundation as partners). For those orgs, today is the day Mythos becomes reachable from the Claude Code CLI without per-call config gymnastics. Everyone else — including most readers of this post — is one approval and a feature flag away from caring.
2. The keynote itself. Dario Amodei opened with three lines that have been doing the rounds on X all morning: that Anthropic believes other major U.S. AI labs are roughly one to three months behind on capabilities; that there is “no issue with putting more compute into Mythos”; and that Anthropic’s revenue grew 80x on an annualized basis as of Q1. Those are positioning statements, not roadmap commitments — but they’re the clearest competitive frame Anthropic has put on the record this year, and they explain why every announcement adjacent to today’s event has been about owning the application layer rather than chasing OpenAI’s model launches.
3. Live demos and workshops. “What’s new in Claude Code” sessions, MCP ecosystem deep-dives, and one-on-one office hours with Anthropic engineers — including Boris Cherny on Claude Code, Ami Vora on product, and Angela Jiang on the API and SDKs. The demos people are screen-recording on X are not new product launches; they’re walkthroughs of the existing stack at full power. A nine-agent Claude Code setup running 500 client tasks per month. A TradingView quant-analyst install with full prompt scaffolding. A two-hour live build of an entire operating system with the GitHub repo dropped publicly. None of these required a new model — they’re showing what’s already possible with what’s been live since Sonnet 4.6 and the auto-mode March release.
What you were told would launch — and when it actually shipped
Five things were on the prediction list. Here’s where they actually are.
Claude Sonnet 4.8. Not launched today. The expectation came from an npm packaging exposure in March that surfaced about 512,000 lines of Claude Code TypeScript, including model identifiers that pointed to a forthcoming Sonnet 4.8 release. Eight hours into the SF event, Anthropic has not posted a release blog and the model-list endpoints on Bedrock and Vertex still return Sonnet 4.6 as the latest in that lane. If you have a migration plan ready, hot-pin Sonnet 4.6 for production traffic and watch the npm registry, the Anthropic news page, and the platform release notes — that’s where it lands first when it lands.
Claude Code “auto mode.” Already GA. Anthropic posted the auto-mode launch on March 23, 2026 and rolled it to research preview on Team plans March 24, with Enterprise and API following in the days after.
--permission-mode plan for routine work like test runs and lint fixes, you’re paying interruption tax for nothing.
Claude Mythos preview. Already out, but only for one tier of customers. Anthropic released claude-mythos-preview on April 7 as a gated research preview through Project Glasswing for defensive cybersecurity workloads with the named partner consortium. What’s new today (small) is that the CLI 2.1.129 gateway-discovery flag makes Mythos addressable from Claude Code without manual configuration for those eligible orgs. If you’re not in Glasswing, this is informational — the model is not yet a public route.
Claude Marketplace partners. Already in market. The marketplace launched March 6 with the six partners — GitLab, Harvey, Lovable, Replit, Rogo, Snowflake — announced as launch partners that day. Anthropic does not take a cut, and enterprises can put marketplace SKUs against existing Anthropic spend commitment with single-vendor invoicing. The IT-procurement story is real; it’s just not new today.
Claude × Microsoft 365 GA. Already documented. Excel, PowerPoint, and Word add-ins shipped on April 30 with the cross-app context-share documented in Anthropic’s support pages, and yesterday’s New York “Briefing: Financial Services” event layered on the ten finance agent templates and the Moody’s MAS native app. The Outlook beta is the only piece still maturing. So if your Q3 stack-planning meeting is tomorrow and someone says “let’s wait for the Microsoft 365 announcement,” you can tell them it already happened — twice — and your decision lives in licensing math, not in launch timing.
The 90-day impact ranking
If your job is to fold any of this into your team’s next quarter, here’s the ranking that survives once you strip out the SF stage lighting.
| What | When it shipped | Q3 dev-team impact |
|---|---|---|
| Auto mode (Claude Code) | March 24, 2026 (research preview) | Move — turn it on for repository hygiene, test runs, lint fixes; gate sensitive paths |
| M365 add-ins (Excel/PPT/Word) | April 30, 2026 | Move if M365-anchored; Watch Outlook beta if Outlook-anchored |
| Claude Marketplace + 6 partners | March 6, 2026 | Watch — depends on whether GitLab/Harvey/Replit/Snowflake/Lovable/Rogo are on your renewal list |
| Mythos preview (gateway discovery in CLI 2.1.129) | April 7, 2026 + today’s CLI bump | Watch — actionable only if you’re a Glasswing org |
| Sonnet 4.8 | Not yet (T-0 expected) | Watch — pin Sonnet 4.6 for prod, run parallel evals when 4.8 lands |
Two things on this list are genuinely Move-this-quarter calls. The other three are watch items unless you’re in a specific cohort. That’s a different shape than the keynote-day energy suggests, and it’s the shape your Slack should have by Thursday.
What this means for you
If you’re an engineering manager already on Claude Code Team or Enterprise. Auto mode is the single highest-leverage feature you’re probably under-using. Turn it on for the repos where the blast radius is contained — repository hygiene, lint cleanup, test suite runs — and gate the production-config paths and migration tooling behind explicit approval. The transcript-classifier safeguard runs on Sonnet 4.6 and operates as a fast filter plus a chain-of-thought second pass; it reduces prompts without skipping all permissions. If you waited on auto mode because of the “research preview” tag, the March release is mature enough that it’s the default recommendation now.
If you’re an IT buyer evaluating Q3 stack decisions. The M365 GA already happened. The decision in front of you is licensing math (Copilot at $30 per seat per month vs Claude Enterprise tiering), data-residency (Microsoft tenant vs Anthropic on AWS/GCP), and SCIM/SIEM parity. None of those changed today. If your Q3 vendor-policy refresh has been waiting for “official confirmation,” you have it — go run the seat-count math.
If you’re inside Project Glasswing. Today’s CLI 2.1.129 is the moment Mythos becomes a one-flag away addition to your Claude Code routing. Set CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1, restart your CLI, and check the model list — claude-mythos-preview should appear if your org’s gating is correctly wired. If it doesn’t, that’s a partnership-config issue to escalate through your Glasswing contact, not a CLI bug.
If you’ve been waiting for Sonnet 4.8 to plan a migration. Today is not the day. Hot-pin Sonnet 4.6 for production traffic, schedule a 24-48 hour parallel-eval window for whenever 4.8 does ship, and watch Anthropic’s news blog, the npm registry for the @anthropic-ai/claude-code package, and the Bedrock/Vertex model-list endpoints — those are the three places the launch will land first. The npm registry is usually the earliest signal because the CLI tags often surface the model identifier before the marketing page goes live.
What today’s news can’t fix
A few honest limits, because the keynote energy hides them.
Auto mode reduces interruption tax; it does not replace review for sensitive operations. The transcript classifier is a substitute for a human approver, not a guarantee. Use it where you trust the general direction and have rollback hygiene, not as a blanket setting on a repo with production-config paths.
The CLI’s gateway model discovery doesn’t grant you access to Mythos — it makes Mythos addressable if your org already has access. The reverse-engineering corner of X has been excited about the claude-mythos-preview identifier all morning, and most of that excitement is misplaced.
The Marketplace doesn’t dissolve your existing vendor contracts. If you’re already on a multi-year GitLab Ultimate or Harvey enterprise contract, marketplace pricing is informational, not a renegotiation lever. Treat it as a procurement-consolidation option for your next renewal, not a refund channel.
The Microsoft 365 cross-app context-share is real but it’s per-session, not org-wide. The handoff from Excel to PowerPoint to Word to Outlook preserves state inside one Claude session; it doesn’t establish a persistent cross-document memory layer. That’s a different product (and one of the things you’ll want to push Anthropic on at office hours if you’re at the SF event).
Sonnet 4.8 didn’t ship today. That doesn’t mean the leak was wrong — it means the timing was wrong. Code with Claude London on May 19 and Code with Claude Tokyo on June 10 are the next two reasonable launch windows; the second one matches Anthropic’s pattern of using regional events to land model releases that didn’t make the SF stage.
The bottom line
Code with Claude SF is more victory lap than launch event. The model-launch surface area Anthropic shipped over the last sixty days — auto mode in March, Mythos preview in April, M365 GA on April 30, the Marketplace expansion in March, the financial-services agent templates yesterday — is what the SF demos are showing off. Today’s actual product news is a CLI bump that exposes one new model identifier behind a feature flag for one specific cohort.
If you came in expecting a Sonnet 4.8 release notes URL by lunch, you’re going to leave disappointed. If you came in to watch a company decide that the application layer is the more interesting battle, you’re getting exactly what was advertised.
The two practical Q3 moves out of all this: turn on auto mode for the repos where it’s appropriate, and run the M365 vs Copilot seat-count math your CFO has been asking for. Both of those decisions stand whether or not Sonnet 4.8 lands tomorrow.
If you’re a developer who wants to actually use what’s already shipping — auto mode, MCP servers, Claude Cowork sessions across multiple files — our Claude Code Mastery course walks through the live setup, the safe patterns for auto mode, and the routine workflows the SF demos are showcasing today. It’s the fastest path from “I should turn this on” to “this is how my team works now.”
Sources
- Code with Claude San Francisco — May 6, 2026
- Code with Claude — Anthropic blog announcement
- Claude Code auto mode launch — Anthropic blog (March 2026)
- Auto mode engineering deep-dive — Anthropic Engineering (March 2026)
- Anthropic gives Claude Code new auto mode — TechRadar
- Claude Marketplace launch — Anthropic platform page
- Anthropic launches Claude Marketplace — VentureBeat (March 6, 2026)
- Claude Marketplace third-party services — SiliconANGLE
- Use Claude for PowerPoint — Anthropic Help Center
- Anthropic deepens push into Wall Street with Microsoft 365 integration — Fortune (May 5)
- Agents for financial services and insurance — Anthropic blog (May 5)
- Claude Sonnet 4.8 release date — NxCode (leak coverage)
- Choose a permission mode — Claude Code Docs
- What’s new — Claude Code Docs
