Anthropic’s August acceleration peaked this week: unified memory across Chat and Cowork, a cybersecurity-only model in an output-only sandbox, five APIs going GA on the same day, an MCP roadmap redesigning agent identity, and Python SDK v1.0.0 with breaking changes — all in seven days. For CTOs and engineering leads, the implications span budget, security posture, migration timelines, and architecture decisions shaping the next 12 months.

Unified Memory Across Chat and Cowork

Anthropic merged Chat and Cowork memory into a single unified layer. Context flows bidirectionally. Memory updates during conversations; users can view, edit, or delete topics in Settings → Memory.

Sensitive topics (health, race, ethnicity, religion, politics, gender identity) are excluded by default with an opt-in toggle. SSNs, government IDs, criminal history, immigration status, and AUP violations are never saved.

Free/Pro/Max: on by default. Team/Enterprise: off by default, admin-controlled. Legacy memory export deadline: September 9, 2026. Local Cowork sessions don’t use memory; Claude Code memory stays separate.

Mythos 5 in Claude Security: The Model That Can’t Write Exploits

Claude Mythos 5 — Anthropic’s most capable, most restricted cybersecurity model — now powers Claude Security for enterprise vulnerability scanning. Users never interact with it directly: no prompt box, no raw access. The model returns scan results only and “cannot be steered into writing exploits.”

Workflow: admin enables Claude Security, user selects a GitHub repo at claude.ai/security, Mythos 5 traces data flows and Git history, returns findings with CWE category, confidence, severity, and suggested fix. Patches open in Claude Code using the org’s existing models. Every patch requires human approval. Billing is standard token usage.

Four tracks: Enterprise scan output (live), partner integration with Accenture/BCG/Deloitte/Infosys/PwC (early), the $35M Defender Advantage Fund for open-source security orgs, and an expanded Cyber Verification Program.

Five APIs Go GA Simultaneously

August 19 was Anthropic’s biggest platform release of the year. Beta headers can now be removed from production code.

Computer use (computer_toolset_20260801) — multi-action turns (click, type, screenshot in one call). 20–40% fewer round trips, ~30% lower cost per workflow. HIPAA-eligible.

Browser use (browser_toolset_20260801) — new toolset, not renamed computer use. Drives a browser viewport via accessibility tree instead of full-desktop screenshots. Both work on Fable 5, Mythos 5, Opus 5, Sonnet 5, Opus 4.8.

Files API (/v1/files) — 5x rate limits (500 RPM), 1 TB per org, file expiration, pagination, ids[] filter. GA response format changed — update parsers.

Agent Skills API (/v1/skills) — 46 client products, 85,000+ public skills. Prerequisites: API key + code execution.

Enterprise Admin API — user management GA across all SDKs. Extended August 26 with full Organization API endpoints (workspaces, API keys, rate limits, cost reports). Auth via API key or org:admin OAuth token.

Legacy Claude API Workbench retired August 17 — no recovery path for unexported prompts.

Python SDK v1.0.0: Real Breaking Changes

The SDK jumped to v1.0.0 on August 20. The HTTP layer moved from httpx to httpx2 — old httpx.Client throws TypeError, and tracing/APM/mocking tools that patch httpx will silently record zero SDK requests. Python floor moved from 3.9 to 3.10. temperature/top_p/top_k removed from messages endpoints. Legacy Text Completions API removed entirely. AnthropicBedrock() now requires a region. compaction_control replaced by server-side context management beta.

Claude Code’s /claude-api upgrade walks codebases through 0.x → 1.x. If you can’t migrate immediately, pin anthropic==0.125.0.

MCP Roadmap: Agent Identity, Progressive Discovery, End of Bearer Tokens

Published August 22 by core maintainers David Soria Parra and Den Delimarsky. Five priorities for 6–12 months:

  1. Agentic messaging — server-initiated events via webhooks/channels; Tasks extension matures toward main spec.
  2. HTTP-native transport — Streamable HTTP as the single binding for all servers, including local.
  3. Agent identity — DPoP (RFC 9449) binds tokens to requesting clients so stolen bearer tokens can’t replay. Workload Identity Federation for AWS IAM, GCP, Azure. Long-lived pasted API keys being designed out.
  4. Progressive discovery — servers reveal tools gradually instead of loading entire catalogs into context. 100-tool catalogs are now a known anti-pattern.
  5. SDK experience — better ergonomics, expedited SEP review.

Scale signal: @modelcontextprotocol/sdk hit 50.7M weekly npm downloads — more than OpenAI (34.5M) and Anthropic (33.9M) SDKs combined. MCP is now default plumbing.

Enterprise-Managed MCP Auth GA

Admins authorize a connector once; employees inherit access via IdP groups — no per-user authorization. First production implementation of MCP’s Enterprise-Managed Authorization extension. Supported: Datadog, Notion, Slack (new), Asana, Atlassian, Canva, Figma, Granola, Linear, Supabase. Coming: Exa, Miro, Zoom. Okta integration via Cross App Access protocol.

Claude Code: 16 Releases in 20 Days

v2.1.243 (60 changes): loops breakdown in /usage, modelPicker for org-approved model lists, modelPricing for contracted rates in cost displays, separate promptCacheTtl/subagentPromptCacheTtl, keyless /login.

v2.1.246 (security): credential scoping fix — API keys for third-party gateways no longer sent to Anthropic in telemetry. Auto mode tab in /permissions. Wildcard Bash allow rule warnings.

v2.1.247: SendFeedback tool, /claude-api cost-optimize for spend profiling and optimization suggestions.

Also: Concise output style (v2.1.237), auto-continue after usage limits (v2.1.234), Alpine/musl support (v2.1.239), Bedrock streaming bug fix that was silently doubling API charges (v2.1.239).

The IPO Looms

Revenue run rate: $65B (end of July). Q2 2026: $11.5B — 14x YoY, overtaking OpenAI’s $6.7B for the first time. Confidential S-1 filed June 1, targeting October Nasdaq listing. Investor target: $2T+. The governance-first strategy (audit logs, compliance API, data retention controls, customer-managed encryption keys) is the differentiator for the public market narrative.

Deadlines

  • Sep 9, 2026 — Legacy memory export deadline
  • Remove GA’d beta headers — files-api-2025-04-14, skills-2025-10-02, ce-user-management-2026-07-13
  • Python SDK v1.0.0 migration — audit httpx patching (silent breakage), removed params, Bedrock region
  • MCP 2026-07-28 migration — identify deprecated features (Roots, Sampling, HTTP+SSE)
  • Sep 29 — Sonnet 4.5 retirement · Oct 15 — Haiku 4.5 retirement · Nov 24 — Opus 4.5 retirement

Action Items

  1. Enable memory for Team/Enterprise in Org Settings → Capabilities. Decide your stance on sensitive-topic memory before users opt in.
  2. Evaluate Claude Security — Enterprise public beta at standard token pricing. Output-only access means no exploit-writing risk.
  3. Migrate Python SDK — the httpx silent breakage is the item most likely to bite you in CI.
  4. Review MCP roadmap — DPoP and Workload Identity Federation will change credential management for agents.
  5. Set up Enterprise-Managed MCP Auth — if you use Okta, admin-authorize-once eliminates per-user friction.
  6. Configure modelPricing and modelPicker — feed contracted rates into cost displays, curate model selection for your org.
  7. Export legacy memory by September 9 — after this date, it’s gone.

The platform is building infrastructure for autonomous, compliant, enterprise-grade AI at a scale anticipating public market scrutiny. Unified memory, a security-only model with no direct access, five GA APIs, a published protocol roadmap, and a major SDK version in one week is not coincidence — it’s a pre-IPO positioning sprint. The question is whether your architecture can absorb the pace.

Follow along at x.com/kkaminsk for weekly analysis of the Claude ecosystem.