KDCube
← AI Industry News

Cheaper Cache, MCP That Acts, and a Firewall for Agents

Anthropic's Claude Fable 5.1 and Mythos 5.1 hold sticker prices but cut cache reads 75% (to $0.25/M), so heavily agentic workloads run roughly 45% cheaper, while Docusign prepares to open its MCP Server to every agent on Sept 30 for governed signing. Meanwhile AIR Security launched with $50M to firewal...

Highlights

  • Anthropic's Claude Fable 5.1 and Mythos 5.1 keep sticker prices flat but cut cache-read cost 75% (to $0.25/M), so a heavily agentic workload runs roughly 45% cheaper — an economics story, not a benchmark one. (Anthropic)
  • Docusign will open its MCP Server to every agent on Sept 30, making agreement drafting, review, and governed signing callable natively from Claude, ChatGPT, Gemini, Copilot, and Slack. (PR Newswire)
  • Alibaba's Qwen3.8-Max-0902 snapshot holds price and 1M context but is re-tuned for coding and multi-tool orchestration, and vendor numbers put it atop the CodeArena front-end leaderboard. (TechNode)
  • AIR Security left stealth with a $50M seed to build an inline "firewall" that vets every skill, plugin, and MCP server across an agent supply chain. (SiliconANGLE)

Key Signals

  1. Cache economics become the real model-pricing lever — Sept 1, official

    Fable/Mythos 5.1 leave input/output rates unchanged ($10/$50 per M) while dropping cache reads from $1.00 to $0.25 per M, with Anthropic estimating ~25% lower cost on typical workloads and ~45% on agentic ones. (Anthropic, VentureBeat) For operators, the win now comes from context reuse and prompt-cache hit rates, not from picking a cheaper tier.

  2. MCP graduates into the enterprise action layer — Sept 4 announcement, Sept 30 open

    Docusign is exposing agreement intelligence and governed signing (via its Iris engine) to any MCP client with account-level admin controls and multi-region infrastructure. (PR Newswire) This moves MCP past read-only connectors into consequential, policy-bound actions — the exact surface enterprise governance teams need to inspect before Sept 30.

  3. The agent supply chain gets its own perimeter — Sept 1

    AIR's launch research flagged 17,800+ public AI add-ons (6.7M installs) pulling from untrusted instruction sources, plus skills impersonating Anthropic and OpenAI to slip past review. (SiliconANGLE) It lands alongside runtime-governance entrants like AccuKnox's AgentZ, which sandboxes each agent with scoped filesystem, network, and credential controls. (GlobeNewswire)

Why It Matters / What To Watch

  1. Re-cost your agent workloads around cache, not tiers.
    • Instrument cache-read hit rates before assuming a model swap saves money; the 5.1 change rewards stable, reused context most. (Anthropic)
    • Model the 5-minute vs 1-hour cache-write premium ($12.50 vs $20/M) against how long your agents actually idle between turns. (VentureBeat)
  2. Treat MCP action servers as governed integrations, not plugins.
    • Inventory which MCP servers your agents can call before Docusign's Sept 30 opening, and confirm admin-level scoping on any that can execute agreements. (PR Newswire)
    • Pair connector adoption with supply-chain vetting — discover, evaluate, and be able to revoke every skill/MCP server, the gap AIR is selling into. (SiliconANGLE)
  3. Watch whether coding-tuned open models pressure the frontier.
    • Qwen3.8-Max-0902 holds price while claiming a top CodeArena result; validate on your own agentic coding evals before trusting the leaderboard delta. (TechNode)

Quick Links