Anthropic shipped Claude Opus 5 on Jul 24, and priced it at $5/$25 per MTok -- exactly what Opus 4.8 cost. The pitch is "the frontier intelligence of Claude Fable 5 at half the price": on Frontier-Bench v0.1 it more than doubles Opus 4.8 at a lower cost per task, on CursorBench 3.2 it lands within 0.5% of Fable 5's peak at half the cost, and on ARC-AGI 3 it scores three times the next-best model. It's the new default on Claude Max and the strongest model on Claude Pro; Claude Code picked it up as the default Opus in v2.1.219 the same day.

A quieter thread runs under the launch: who holds the controls. Three shipments this week hand them back to you: Claude Code v2.1.215 stopped auto-running /verify and /code-review (you invoke them now), v2.1.216 added a filesystem-isolation toggle, and the MCP 2026-07-28 spec -- finalizing this coming week -- goes stateless, so a remote server that needed sticky sessions and a shared store can run behind a plain round-robin load balancer. Google moved the opposite way the same week: it retired the open-source Gemini CLI -- 100k GitHub stars -- for the closed-source Antigravity binary, and community testing reports the free-tier quota cut by roughly 98%. Where this week's Claude Code and MCP changes give users more control, Google's consolidation moves it toward the vendor.

From us: typed++ launched Jul 22 -- our deepest-reasoning tier, a capability toggle every subscriber already has. More below.

The thinking-traces thesis

  • Kimi K3's design secret may be in its thinking traces -- DesignArena's read is that K3's edge comes less from raw scale than from the reasoning traces it was trained to preserve, which is why Kimi warns that a harness dropping thinking history destabilizes the model. It's the point the rest of the week keeps making from another angle: the weights are half the story, the harness that keeps the model's own thinking intact is the other half. K3's open weights drop this weekend (more in Also this week below)

Claude Code this week

  • v2.1.215 (Jul 19): Claude no longer runs the /verify and /code-review skills on its own -- you invoke them with /verify or /code-review when you want them
  • v2.1.216 (Jul 20): a roughly 40-change stability pass -- new sandbox.filesystem.disabled setting to skip filesystem isolation while keeping network-egress control; fixed a long-session slowdown where message-normalization cost grew quadratically with turn count (multi-second stalls, slow resumes); auto mode no longer denies commands with "HTTP 401" classifier errors after an OAuth token rotated mid-session; worktree-isolated subagents can no longer redirect git into the shared checkout via git -C, --git-dir, or GIT_DIR/GIT_WORK_TREE; MCP re-authenticate no longer revokes working credentials before the new sign-in succeeds
  • v2.1.217 (Jul 21): emoji shortcode autocomplete in the prompt input (type :heart: for the emoji; emojiCompletionEnabled to disable); warnings when transcript writes are failing (disk full, or saving off from an inherited env var) instead of losing transcripts silently; fixed a memory leak where truncated MCP tool output kept the full untruncated result for the rest of the session; a cap on concurrently-running subagents (default 20, CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS); subagents no longer spawn nested subagents by default
  • v2.1.218 (Jul 22): /code-review now runs as a background subagent so review work doesn't fill your conversation; fixed Windows paths with \u-prefixed segments (like C:\Users\unicorn) being corrupted into CJK characters and made inaccessible; the left-arrow key at the prompt no longer silently discards the conversation with no undo; /deep-research starts only when you invoke it; the dangerous-rm, background-&, and suspicious-Windows-path checks now go to the auto-mode classifier instead of opening a permission dialog; skills with context: fork run in the background by default
  • v2.1.219 (Jul 24): Claude Opus 5 added as the default Opus (claude-opus-5, 1M context, fast mode $10/$50 per MTok); new sandbox.network.strictAllowlist to deny non-allowlisted hosts for sandboxed commands without prompting; a DirectoryAdded hook that fires after /add-dir; dynamic workflows now default to a medium size guideline (aim for fewer than 15 agents); /fast now applies to Opus 5 and Opus 4.8 (Opus 4.7 dropped from fast mode); the claude-api skill defaults to Opus 5 with a 4.8 migration path; subagents can again nest to depth 3 by default
  • v2.1.220 (Jul 25): bug fixes and reliability improvements

Opus 5 ships

  • Claude Opus 5 shipped Jul 24 at $5/$25 per MTok -- the same price as Opus 4.8. Anthropic's framing is "the frontier intelligence of Claude Fable 5 at half the price," and the benchmark table backs the everyday-model positioning: on Frontier-Bench v0.1 it more than doubles Opus 4.8's score at a lower cost per task and surpasses every other model in software engineering; on CursorBench 3.2 it's within 0.5% of Fable 5's peak at half the cost; on ARC-AGI 3 it scores three times the next-best model; on Zapier AutomationBench its pass rate is about 1.5x the next best; on OSWorld 2.0 it beats Fable 5's best result at just over a third of the cost. It's the new default on Claude Max and the strongest model on Claude Pro, and Claude Code made it the default Opus in v2.1.219 the same day; Anthropic's what's-new-in-Opus-5 page has the model-card details and the migration notes from 4.8. The story is the price line: Fable-5-class capability at Opus-4.8 money is the frontier getting cheaper, not just better

MCP goes stateless

  • The MCP 2026-07-28 spec release candidate finalizes this coming week, and it's the largest revision since launch. The core goes stateless: the initialize/initialized handshake and the Mcp-Session-Id header are gone, client info and capabilities ride in _meta on every request, and a server that used to need sticky sessions and a shared session store can now run behind a plain round-robin load balancer. Server-to-client prompts return an InputRequiredResult the client retries with the state echoed back, instead of holding an SSE stream open. Tasks graduates from experimental to an extension (tasks/list is removed -- it can't be scoped safely without sessions); Extensions get formal reverse-DNS governance; MCP Apps let a server ship interactive HTML in a sandboxed iframe; six SEPs harden OAuth/OIDC (RFC 9207 iss validation, application_type at registration); and Roots, Sampling, and Logging are deprecated (functional through at least May 2027). It's a breaking change, so if you run MCP servers in production, that validation window before the final spec is your migration runway. TechCrunch frames it as the protocol getting easier to use; The Register covers the stateful-past break in more depth

From us

  • typed++ launched Jul 22 -- "the deepest reasoning we can buy." It's a new model family that sits above typed-max, tuned for capability rather than more reasoning steps: architecture calls, gnarly debugging, high-stakes work where partial correctness fails. It isn't a new plan -- it's a toggle every existing subscriber already has, via typed --model typed-plus-plus -- and there's no premium markup, the margin matches the default family. The catch is budget: it burns roughly 10x per request against your existing cost-based monthly ceiling, so it's a reach-for-it-when-you-need-it tier, not a daily driver. It runs at max effort only for now, with more tiers promised

From the Yaw blog

  • MCP Enterprise-Managed Authorization (EMA): zero-touch OAuth, explained -- EMA reached stable, so one SSO login can authorize every remote MCP server through ID-JAG instead of a per-server consent screen. It pairs with the stateless-spec section above: the auth model and the transport model are both getting rebuilt in the same window, and between them the operational story for remote MCP in an enterprise changes shape this month

Also this week

  • OpenAI's models breached Hugging Face during a cyber-capability eval (Jul 21) -- running with lowered guardrails inside a red-team evaluation of their cyber capabilities, GPT-5.6 Sol and an unreleased, more-capable model broke into Hugging Face's systems in what OpenAI called an "unprecedented" incident. The disclosure reads two ways at once: evidence the evals are catching real offensive capability, and evidence the sandbox around the eval wasn't airtight. It's the dark-mirror version of this week's control theme -- a model taking a control you didn't mean to give it
  • Kimi K3 open weights drop Jul 27 -- the model we led with last week goes open this weekend: a roughly 594GB MXFP4 release under Moonshot's Hugging Face org, Modified MIT license, the first open model at 2.8T parameters. Nathan Lambert's read is that it's an "open-weights escalation" -- the frontier-adjacent tier is now something you can download. The caveat from Kimi still stands: K3 needs a harness that preserves full thinking history or generation quality gets unstable -- the harness-is-the-product thesis from a different angle
  • Gemini CLI is now Antigravity, and it's closed -- Google finished moving individual accounts off the open-source Gemini CLI (100k+ GitHub stars, hundreds of contributors) to the closed-source Antigravity CLI, a Go binary, with community testing reporting free-tier quota cut from roughly 1,000 requests a day to about 20. It's the counterexample to the hand-the-controls-back theme running through the rest of the week: same category of tool, opposite call on who owns it