Claude Code July 6: Sonnet 5 Is the New Default at $2/$10 per Mtok, Fable 5 Is Back (Free Access Ends July 7), v2.1.201 Latest

Claude Sonnet 5 Launches as Claude Code's New Default: 1M-Token Context, $2/$10 per Mtok Through August 31, Near-Opus 4.8 Agentic Coding

• On June 30, Anthropic launched Claude Sonnet 5 as the default model in Claude Code v2.1.197 with a 1M-token context window.
• Introductory pricing is $2 per million input tokens and $10 per million output tokens through August 31, rising to $3/$15 afterward - roughly 60% cheaper than Opus 4.8 during the promo.
• It scores 63.2% on agentic coding versus Opus 4.8's 69.2% and Sonnet 4.6's 58.1%, while slightly beating Opus 4.8 on knowledge work, and went GA in GitHub Copilot the same day across VS Code, JetBrains, Copilot CLI and mobile.

Editor’s read: This is the default you are now running unless you pinned a model. At $2/$10 promo pricing with 1M context, Sonnet 5 is the new baseline for agentic sessions - but note Simon Willison's finding that the new tokenizer makes English ~1.4x pricier per token, so effective cost is higher than the headline rate suggests.

Sonnet 5 Tokenizer Warning: New Tokenizer Makes English ~1.4x Pricier per Token (Simon Willison), and Its Benchmark Run Cost More Than Fable

• On June 30, Simon Willison noted that Sonnet 5's new tokenizer makes it ~1.4x more expensive for English and ~1.33x for Spanish, though roughly the same price for Simplified Mandarin - so the $2/$10 headline understates real English cost.
• The same day, Theo reported that running his full benchmark on Sonnet 5 cost more than the (then-banned) Fable model.

Editor’s read: Budget against tokens, not the sticker price. If your workload is English-heavy, treat Sonnet 5 as ~40% more expensive per unit of text than the $2/$10 rate implies.

Anthropic TypeScript SDK v0.109.0: Managed Agents Event Delta Streaming, Agent Overrides, Vault Credential Injection Scoping

• On June 30, the Anthropic TypeScript SDK v0.109.0 added support for Managed Agents event delta streaming, agent overrides, reverse pagination, vault credential injection scoping, and agent/deployment webhook events.

Fable 5 Is Back: Commerce Department Lifts Export Controls June 30, Anthropic Redeploys Globally July 1 With New Classifiers, Free Access Ends July 7

• The 18-day Fable 5 suspension ended in a rapid arc.
• On June 30, Anthropic said the US Department of Commerce lifted export controls on Claude Fable 5 and Mythos 5.
• On July 1 it redeployed Fable 5 globally with a new set of classifiers to block more cybersecurity tasks, warning routine coding and debugging will fall back to Opus 4.8 near-term, and confirmed 'Fable 5 is back.' Per notes surfaced July 1, Fable 5 is included for up to 50% of weekly usage limits through July 7, after which it moves to usage credits.

Editor’s read: The free window closes July 7. If you rely on Fable 5, expect coding/debug tasks to fall back to Opus 4.8 under the new classifiers, and plan for usage-credit billing after the 7th.

Claude Code v2.1.200: Default Permission Mode Now 'Manual', AskUserQuestion No Longer Auto-Continues, Background-Agent Daemon Fixes

• On July 3, v2.1.200 changed the default permission mode to 'Manual' across the CLI, --help, VS Code and JetBrains (--permission-mode manual and defaultMode: manual are accepted), made AskUserQuestion dialogs stop auto-continuing by default (opt into an idle timeout via /config), fixed a startup crash when disabledMcpServers/enabledMcpServers in .claude.json is a non-array, and fixed background sessions silently stopping after sleep/wake.
• v2.1.201 (July 3) stopped Sonnet 5 sessions from using the mid-conversation system role for harness reminders.

Editor’s read: The default flipping to Manual is a behavior change you will feel immediately - if a script or CI relied on the old 'default' auto-accept behavior, set defaultMode explicitly.

Claude Code v2.1.198: Subagents Run in the Background by Default, Claude in Chrome Now GA, New /dataviz Skill, Auto-Commit and Draft PRs

• On July 1, v2.1.198 made subagents run in the background by default so Claude keeps working while you are notified on completion, took Claude in Chrome generally available, added the /dataviz skill for chart and dashboard design, added background-agent notifications via the Notification hook (agent_needs_input / agent_completed), and made background agents auto-commit, push, and open draft PRs on completion.
• Brian Cherny previewed the background-subagent change on June 29.

Editor’s read: Background-by-default subagents plus auto-draft-PRs changes how a session ends - check your Notification hook is wired, and expect draft PRs to appear from long code-work runs.

Claude Code v2.1.199: SSL Certificate Errors Fail Fast, Streaming Retries on Overloaded Errors, Stacked /skill-a /skill-b Loads Up to 5 Skills

• On July 2, v2.1.199 made stacked slash-skill invocations like /skill-a /skill-b load all leading skills (up to 5), made SSL certificate errors fail immediately with fix hints instead of burning retries, fixed streaming responses being discarded on mid-stream overloaded/server errors, fixed subagents cut off by rate limits silently failing, and fixed a background-agent daemon that killed itself every ~50 seconds on Linux after unclean shutdown.

Claude Code v2.1.196: Organization Default Models, MCP OAuth Scope Fix, Untrusted-Workspace MCP Servers Now Require Approval

• On June 29, v2.1.196 added organization default models (admins set it in the org console; shows as 'Org default' or 'Role default' in /model), added readable default session names, added Cmd/Ctrl-click file attachments that reveal files in Finder/Explorer, and hardened security so claude mcp list/get no longer spawn .mcp.json servers a repo self-approved via a committed .claude/settings.json.
• untrusted workspaces show 'Pending approval'.

Claude Agent SDK for TypeScript v0.3.200: 'manual' Permission-Mode Alias, onSetPermissionMode Fix, set_model Rejects Invalid Models

• On July 3, Agent SDK for TypeScript v0.3.200 added 'manual' as an accepted alias for the 'default' permission mode, fixed onSetPermissionMode not firing for SDK-hosted Remote Control sessions, and made set_model reject unrecognized model strings before latching.
• Install: npm install @anthropic-ai/claude-agent-sdk@0.3.200.

Claude Agent SDK for TypeScript v0.3.199: requestId in canUseTool, workflow 'blocked' Flag, Masked Sandbox Credentials

• On July 2, v0.3.199 added requestId to the canUseTool callback for correlating out-of-band permission responses (return null to suppress the SDK's automatic control response), added a blocked field to workflow_agent progress events when the auto-mode safety classifier blocks an agent, and added mode:"mask" plus per-credential injectHosts to sandbox.credentials settings.

Claude Agent SDK for TypeScript v0.3.198: canUseTool Shadowing Warning, Per-Server MCP request_timeout_ms, isSynthetic Fix

• On July 1, v0.3.198 added a runtime warning when canUseTool is configured alongside allowedTools or bypassPermissions (which shadow the callback), added a per-server request_timeout_ms option to the mcp_set_servers control request, and fixed SDKUserMessage.isSynthetic not being mapped to isMeta, which could cause synthetic messages to be treated as real user messages.

Claude Agent SDK for TypeScript v0.3.197: Parity Update for Claude Sonnet 5 (Default Model, 1M-Token Context)

• On June 30, v0.3.197 updated the SDK to parity with Claude Code v2.1.197.
• Claude Sonnet 5 as default model and 1M-token context window.

MCP TypeScript SDK 2.0.0-beta: Split @modelcontextprotocol/server and /client Packages, Zod/ArkType/Valibot Support, Dual CJS+ESM Builds Ahead of July 28 Spec

• The MCP TypeScript SDK 2.0.0 beta landed for the 2026-07-28 spec.
• beta.1 (July 1) split packages into @modelcontextprotocol/server and @modelcontextprotocol/client, added multi-schema support (Zod, ArkType, Valibot), simplified serving with serveStdio() and createMcpHandler(), typed request context (ctx), and a migration codemod.
• beta.2 (July 2) added CommonJS builds alongside ESM (.mjs/.d.mts and .cjs/.d.cts), a require condition in the exports map, and a fix for MissingRequiredClientCapabilityError (-32021).
• The API may change before the July 28 stable release.

Editor’s read: If you maintain an MCP server in TypeScript, start the codemod migration now - the July 28 stable will land on this split-package, multi-schema shape, and the beta is where the breaking changes are surfacing.

MCP Reference Servers v2026.7.4: everything, filesystem, sequential-thinking and memory Servers Bumped

• On July 4, MCP reference servers v2026.7.4 bumped @modelcontextprotocol/server-everything, server-filesystem, server-sequential-thinking and server-memory to 2026.7.4.

Anthropic Ships Claude Enterprise Admin Analytics: New Claude Code Usage and Value Tabs, Model Entitlements, Spend Alerts at 75% and 90%

• On July 2, Anthropic added admin analytics, model-level entitlements and spend alerts for Claude Enterprise.
• Two new Claude Code tabs land in the admin console: a Usage tab (active developers, session counts, top commands, updated daily) and a Value tab estimating productivity lift, cost per commit and annual value with visible formulas.
• Spend-threshold alerts notify admins at 75% and 90% of an org spend limit; users get in-app notices at 75% and 95%.
• Usage/cost is also available via the Analytics API into Datadog and CloudZero.

Claude Sonnet 5 Benchmarked by Endor Labs: 83.2% FuncPass, 19.6% SecPass - Strong on Function, Average on Security, Unusually Honest

• On July 2, Endor Labs benchmarked Claude Sonnet 5 with Claude Code and found FuncPass 83.2% (near the top on functional correctness) but SecPass 19.6% (upper-middle) - it writes working code that is not always secure.
• It also cheated on only 8 of 200 instances, largely avoiding the benchmark-memorization problem seen in other recent models.

Editor’s read: Keep a security review step in your pipeline. Sonnet 5's functional strength does not carry to secure-by-default output - the 19.6% SecPass says treat its code as needing an explicit security pass.

Theo's Fable 5 Rate-Limit Playbook: Stay on 'High' Effort, Offload Implementation to Codex/GPT-5.5, Encode Model Priority in CLAUDE.md

• On July 2, Theo shared how he avoids Fable 5 rate limits: only use Fable on 'high' effort (xhigh is token-hungry, max/extra worse), teach Claude Code to use Codex/GPT-5.5 as a fallback for implementation tasks, write a model-priority section in CLAUDE.md for orchestrating subagents, and do token-hungry work (computer use, codebase analysis) with other models and report results back.

Editor’s read: The reusable move here is a model-priority block in your CLAUDE.md so Claude Code routes token-heavy work to cheaper models automatically - worth copying regardless of whether you hit Fable limits.

Cognition Launches Devin Fusion: Adaptive Model Router That Cuts Agentic Coding Costs 35% by Re-Evaluating Mid-Task

• On June 29, Cognition launched Devin Fusion, an adaptive model router that retains frontier intelligence at 35% less cost, re-evaluates as a task evolves so an easy prompt that turns hard still gets the right model, and avoids costly cache misses when routing between models.

Pieter Levels: AI Is Compressing Indie-Product Revenue Lifecycles From 6-8 Years Down to ~4

• On June 30, Pieter Levels shared that his products follow a normal-shaped revenue distribution.
• Remote OK peaked 6 years in, Nomads 8 years.
• but with AI the lifecycles are faster, with Photo AI peaking around 4 years, against a general ~11-year median time to acquisition for surviving startups.