STEVEN STEINBECK
← ALL DISPATCHESDISPATCH 67

WEEK OF APR 4 – 10, 2026

Anthropic walls off its most capable model, and flat-rate agent pricing breaks

A week of hard lines: third-party agent harnesses cut out of Claude subscriptions, a frontier cyber model deliberately withheld, and a federal appeals court letting the Pentagon's blacklist stand.

Anthropic published Claude Mythos Preview, a cybersecurity-capable frontier model it refuses to release publicly, funneling it instead to 40+ critical-infrastructure defenders under Project Glasswing.

It also cut third-party agent harnesses like OpenClaw out of Claude subscriptions — the moment the flat-rate economics of running agents on a consumer plan broke.

WED · Apr 8, 2026modelsmetamultimodalreasoningrl

Meta Superintelligence Labs ships its debut model, Muse Spark — closed-weight, and not the Llama playbook

Meta introduced Muse Spark, the first model out of Meta Superintelligence Labs and the first in its Muse family: a natively multimodal reasoning model with tool use, visual chain-of-thought, and a Contemplating mode that orchestrates multiple parallel reasoning agents.

Unlike the Llama line it is proprietary, rolling out through the Meta AI app, meta.ai, and Ray-Ban and Oakley glasses. Meta describes RL training that maximizes correctness subject to a thinking-time penalty, producing thought compression on evaluations like AIME. Published numbers show it competitive but not dominant: 75.1 on Terminal-Bench 2.0 and 93.5 on IPhO 2025 Theory, but 42.5 on ARC-AGI-2 against Gemini 3.1 Pro's 76.5.

WHY IT MATTERS

A reorganized lab's first output, and a strategy reversal: Meta's frontier models are now closed-weight product differentiators, with reasoning-token efficiency rather than raw scale as the headline claim.

Meta AI — Introducing Muse SparkFortune — Meta unveils Muse Spark
WED · Apr 8, 2026policylitigationanthropicpentagonregulation

D.C. Circuit lets the Pentagon's supply-chain-risk designation of Anthropic stand

A three-judge panel of the US Court of Appeals for the D.C. Circuit denied Anthropic's emergency motion to stay the Department of War's supply-chain risk designation pending appeal, allowing the Pentagon to bar contractors from using Claude on defense work while litigation continues.

The panel wrote that the equities tip in favor of the government, while acknowledging Anthropic raises substantial challenges and will likely suffer some degree of irreparable harm, and expedited oral arguments to May 19. The ruling sits in tension with the Northern District of California's March 26 preliminary injunction, which blocked enforcement of the designation for non-Pentagon agencies and found evidence of First Amendment retaliation.

WHY IT MATTERS

A live test of how far the executive branch can use national-security procurement authority against a model vendor's usage policies — with two federal courts pointing opposite directions, the outcome shapes what safety guardrails frontier labs can afford to keep.

NTD — Appeals court allows Pentagon designation
TUE · Apr 7, 2026modelsfrontiersecurityevalsrelease-policy

Anthropic reveals Claude Mythos Preview and refuses to release it, routing it to defenders via Project Glasswing

Anthropic disclosed a new frontier model, Claude Mythos Preview, and announced it will not be publicly released — access goes to launch partners including AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, plus 40+ additional organizations maintaining critical software.

Anthropic committed up to $100M in Mythos Preview usage credits and $4M in donations to open-source security organizations. Reported capability jumps over Opus 4.6 include 83.1% versus 66.6% on CyberGym, 93.9% versus 80.8% on SWE-bench Verified, and, on Firefox JavaScript-engine exploit development, 181 successes against 2 for its predecessor.

WHY IT MATTERS

The first asymmetric release of a frontier model at this capability tier — the eval numbers that justify withholding it, especially the exploit-generation leap, are the most concrete capability-we-chose-not-to-ship evidence the field has produced.

Anthropic — Project Glasswing
MON · Apr 6, 2026infrastructurehardwaretpusanthropicbroadcom

Anthropic signs multi-gigawatt TPU deal with Google and Broadcom as run-rate revenue passes $30B

Anthropic announced a new agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity expected to come online starting in 2027, most of it sited in the US as an extension of its $50B American infrastructure commitment.

Broadcom's SEC filing put the Anthropic portion at roughly 3.5 GW of TPU-based capacity, plus a long-term agreement to develop and supply custom TPUs for Google and a supply-assurance deal for networking components through 2031. Anthropic said its run-rate revenue has passed $30B, up from about $9B at the end of 2025, with over 1,000 business customers spending more than $1M annually.

Anthropic signs multi-gigawatt TPU deal with Google and Broadcom as run-rate revenue passes $30B
Anthropic

WHY IT MATTERS

The clearest datapoint of the week that frontier AI is now a power-and-silicon procurement race: the same lab that just withheld a model is contracting gigawatts of custom accelerators years out, and its revenue tripled in a single quarter.

Anthropic — Expanded partnership with Google and BroadcomCNBC — Broadcom agrees to expanded chip deals
MON · Apr 6, 2026agentsagent-harnessawsdevopssecurity

AWS ships its first two 'frontier agents' — DevOps Agent and Security Agent — to general availability

AWS moved AWS DevOps Agent and AWS Security Agent from preview to general availability, describing them as frontier agents: systems that work independently across many steps, scale across a portfolio, and run persistently for hours or days.

DevOps Agent correlates telemetry, code and deployment data to investigate incidents across AWS, multicloud and on-prem, integrating with CloudWatch, Datadog, Dynatrace, New Relic, Splunk, Grafana, GitHub, GitLab and Azure DevOps; AWS reports preview customers saw up to 75% lower mean time to resolution, 80% faster investigations and 94% root-cause accuracy. Security Agent does continuous, context-aware penetration testing inside development pipelines. Both are consumption-priced, roughly $30/hour for SRE work and $50/hour for pentesting.

WHY IT MATTERS

Long-horizon, tool-using agents crossed from demo to a priced, supported product on the biggest cloud — and per-hour pricing makes the make-versus-buy comparison against a human SRE explicit.

AWS — DevOps Agent & Security Agent GA
SAT · Apr 4, 2026agentsagent-harnesspricingopenclawanthropic

Anthropic cuts third-party agent harnesses out of Claude subscriptions

Effective April 4 at 12pm PT, Claude Pro, Max and Team subscription limits no longer cover third-party harnesses — enforcement started with OpenClaw and was scheduled to extend to others. Users keep access but must move to Anthropic's pay-as-you-go Extra Usage billing or standard API keys, with a one-time transition credit equal to one month's plan.

Boris Cherny, who leads Claude Code, framed it as a capacity decision: third-party harness traffic patterns were never accounted for in flat-rate plans, and Anthropic cited per-day OpenClaw consumption in the hundreds of dollars at API rates. OpenClaw creator Peter Steinberger argued the timing matched Anthropic copying popular features into its own closed harness.

Anthropic cuts third-party agent harnesses out of Claude subscriptions
TechCrunch

WHY IT MATTERS

If you build on a subscription-plan harness rather than API keys, this is the moment the flat-rate agent economics broke — and the clearest signal that providers will meter third-party agent traffic separately from first-party product usage.

TechCrunchOpenClaw — Anthropic just cut off Claude subscriptions