The Scientist

What Happened

Anthropic shipped three changes that interact badly for anyone running Claude in production. First, subscriptions now convert to dollar-matched API credits across Agent SDK, claude-p, GitHub Actions, and third-party harnesses. The implicit 70-90% subsidy on programmatic usage is gone. Second, Dario Amodei conceded the company planned for 10x growth and got 80x, which is the read on weeks of degraded Claude Code performance: a capacity miss, not a model regression. Third, starting June 15, third-party tool usage (Zed, Conductor, OpenCode, T3 Code) moves to a separate credit bucket with no rollover and overflow at API rates.

The capacity patch is leasing xAI's entire Colossus 1 cluster — 220,000+ GPUs spanning H100, H200, and GB200 — from a CEO who publicly insulted Anthropic three months ago. Rate limits are being raised in parallel: Claude Code 5-hour caps doubling, peak-hours throttling removed, Opus API limits "substantially raised."

Why Sources Disagree on Market Position

Seven independent sources reported Anthropic overtaking OpenAI in enterprise adoption (34.4% vs 32.3% per Ramp). The thing this doesn't tell you is what the metric measures: credit-card billing share, not token volume, not workload criticality, not large-enterprise invoiced spend. OpenAI's point that $1M+ ACV accounts pay by ACH, not card, is correct and material. Ramp's own economist separately flagged Opus 4.7 tripling image costs and mounting reliability complaints. The crossover is real for bottoms-up developer adoption. It is not yet a statement about the Fortune 500.

ServiceNow burned its full-year Claude budget by May. National Life Group's CIO says Claude is 'not great for companies' wanting per-user monitoring. The frontier model most teams build on has no native cost attribution, no SLAs, and no per-user telemetry.

The Compound Effect

Token consumption in agentic workflows is non-linear. A reflection loop or tool-use chain can 10x spend per task with no proportional quality gain, and the per-task variance is wide enough that an average tells you very little. Remove the subscription subsidy on the same date the vendor still ships no per-tenant attribution, and the failure mode is a silent budget overrun that surfaces in the monthly invoice rather than the observability dashboard. Teams with gateway-level logging in place before June 15 absorb the change. Teams discovering it in the invoice do not.

The Production Reality

Vercel's AI Gateway production index is the only multi-tenant usage dataset with 200K+ teams and 7 months of data, which is the disclaimer up front. It puts agentic workloads at 59% of all token volume. Six months ago that number was under 20%. The composition shift is the fastest since completion-to-chat endpoints, and it happened while most eval harnesses were still scoring single-turn responses against reference answers.

Three structural mismatches follow. First, cost models are off by ~5x: agentic traces run 15:1 input-to-output versus the 3:1 most forecasts assume, with heavy cache reuse on some providers and none on others. The thing this doesn't tell you is which providers, so the 5x is a population estimate, not a per-vendor one. Second, eval harnesses measure the minority: single-turn accuracy at 90%+ hides the planner that burns 40K tokens arguing with itself before giving up. Third, the spend/volume split is the routing signal: Anthropic captures 61% of dollars via Opus (reasoning), Google captures 38% of tokens via Flash (throughput), with no vendor loyalty observed.

The Architecture That Works at Scale

Abridge's disclosed architecture across 80M+ clinical conversations is a second data point on the same pattern: cheap fast model triages, expensive model reasons only when needed, with memory externalized from weights and LLM judges calibrated against human annotators. Two independent sources converging on confidence-gated routing across a constellation of models is enough to call it the de facto production pattern, not a research proposal.

If 59% of your tokens are agentic but 100% of your evals are single-turn, you're flying instruments-out — update the harness before you update the model.

What the MCP Overhead Tells You

Glean reports off-the-shelf MCP burning 30% more tokens than retrieval-tuned knowledge graphs on agentic tasks, losing 2.5x head-to-head on preference. The methodology is vendor-published with no disclosure. Treat the magnitude as a hypothesis. The directional claim, that naive tool listings balloon context windows while reranked snippets do the same work cheaper, matches what SAP and ServiceNow concluded independently this week. All three converge on Knowledge Graph grounding + MCP-governed execution as the enterprise agent reference architecture.

Multi-agent decomposition adds a dimension worth a separate ablation. Microsoft's MDASH (100+ agents) beat Anthropic's Mythos on CyberGym by decomposing tasks into scan → adversarial debate → PoC exploitation stages. The 100-agent count is a design choice. No ablation isolates the staging from the ensemble size, so it is possible but not established that the staging is doing the work. The decompose-debate-verify pattern is consistent with ensemble priors and is the cheap experiment to run on any workload with auto-verifiable outputs.

The Threshold Crossing

The UK AI Security Institute evaluated the newest Anthropic Mythos and OpenAI GPT-5.5-cyber on autonomous cyber-offense tasks. Both completed full network takeovers in controlled environments, one capability tier above the prior Mythos generation, which topped out at "advanced persistence." Mythos cleared both of AISI's hardest tests. GPT-5.5-cyber cleared one. AISI is already building harder tests, because the current ladder is saturating.

This is the first time a national evaluator has publicly stated that frontier models can complete an end-to-end attack chain. Neither lab is releasing these variants broadly. Anthropic gates Mythos to select enterprises and government agencies.

The 271:1 Harness Signal

Two teams ran Claude Mythos against C codebases in the same month. Mozilla wrapped a custom agentic harness around existing fuzzing infrastructure and surfaced 271 bugs in Firefox 150, including sandbox escapes, use-after-frees, and race conditions. Daniel Stenberg pointed Mythos at curl with a generic scanner and got exactly 1 low-severity CVE with 4 false positives. His verdict: "primarily marketing."

Same model, two orders of magnitude apart. The variable that moved was the harness. Mozilla's wrapper emits reproducible test cases, scales across ephemeral VMs, and integrates with their security lifecycle. The thing the leaderboard score doesn't tell you is which scaffold the model was wearing. This is the strongest public evidence to date that eval budget dominates model choice by at least 50x on real codebases.

Vulnerability discovery just moved from human-weeks to model-minutes. If the patch SLA is not benchmarked against inference time, the defense is tuned to last year's threat model.

Implications for Agent Deployments

Refusal-rate harnesses calibrated on GPT-4-era capability assumptions will produce false negatives against Mythos-class attackers. The miscalibration is structural, not a tuning fix. For anything agentic with tool access, the eval needs a staged rubric covering recon, initial access, lateral movement, persistence, and exfil, not input-side prompt filters scored in isolation. Microsoft's MDASH shipped 16 real Windows fixes off multi-model bug-hunting, which is the load-bearing data point that the capability has crossed the utility threshold for both offense and defense at the same time.

Three recipe and curation results worth pricing in

This week's drops share a through-line for teams running their own training or distillation: the marginal dollar in model development has moved from raw compute to recipe and curation. Each result shifts unit economics in a direction that matters, and each carries its own replication risk profile.

1. Nous Token Superposition Training (TST)

Reports 2-3x wall-clock speedup at matched FLOPs with no inference-time architecture change, validated from 270M through 10B-A1B MoE. The thing this measures is pretraining throughput. The thing it doesn't measure is whether the speedup holds at frontier scale under independent replication. Single source, but the mechanistic claim is clean. If it holds, it's a free 2-3x on training runs without touching the serving stack.

2. NVIDIA Star Elastic

Claims one post-training run produces a family of reasoning model sizes at 360x lower cost than pretraining a family, 7x better than SOTA compression. Headline numbers of that size always shrink under independent eval. The question is the floor, not the ceiling. A 30x hold would still restructure how size tiers are produced for routing. A 10x hold probably would not.

3. Datology VLM Curation

Hit +11.7 points on 20 VLM benchmarks at 2B params, beating InternVL3.5-2B by about 10 points at 17x less training compute. Produced a near-frontier 4B model at 3.3x lower response FLOPs than Qwen3-VL-4B. The lever was purely data curation, not architecture or scale. This is the clearest evidence this year that curation dominates compute at the VLM frontier.

TST is the one to spike first: it's a training recipe with no inference-side tax. If it replicates at even 1.6x, it pays for itself on the next full run.

Adjacent: SWE-ZERO-12M-Trajectories

Kevin Li released 112B tokens, 12M trajectories, 122K PRs across 3K repos in 16 languages, positioned as the largest open agentic trace corpus. Useful for SFT, reward-model training, and synthetic eval generation. Open releases at this size tend to acquire licensing friction within months. Worth mirroring before that happens.