The Scientist

What Changed This Week

Anthropic tightened the pricing surface this week, and the budget written in October no longer covers the workload run in November. Claude subscriptions now convert to dollar-matched API credits across every programmatic surface: Agent SDK, claude-p, GitHub Actions, and third-party harnesses. The 70-90% effective discount power users were extracting on alternative harnesses is gone. Starting June 15, third-party tool usage (Zed, Conductor, OpenCode, T3 Code) lands in a separate credit bucket with no rollover and overflow at API rates. Dario Amodei has admitted Anthropic planned for 10x growth and hit 80x, which is why they are emergency-leasing xAI's 220,000+ GPU Colossus 1 cluster (H100, H200, GB200). The capacity scramble is the cause. The metering is the consequence.

The Contradiction Worth Surfacing

Sources disagree on what the Ramp 34.4% vs 32.3% crossover means. Several cite it as evidence Anthropic is winning. Others correctly note that Ramp measures who gets billed on a corporate card, which doesn't capture token volume, workload criticality, or production dependency. OpenAI counters that large enterprise contracts go through invoice and ACH, not cards. The directional signal is robust across sources: multi-vendor procurement is now the default. The specific ranking is noise within measurement error.

ServiceNow burned its full-year Claude budget by May. The cost-attribution gap bites most teams within one quarter of going live.

The No-SLA Problem

Anthropic provides no native per-user usage telemetry, no tool-level consumption breakdown, no SLAs on latency or availability, no budget alerts, and no anomaly detection. For enterprise SaaS at this price point, that is anomalous. The thing this doesn't tell you from the pricing page is that you cannot attribute which tenant, prompt, or feature drove the bill until the invoice arrives.

The OpenAI Counter

Sam Altman dropped a 2-month-free Codex enterprise switch promo the same day Anthropic metered. That is a zero-cost evaluation window. The right read is to run it through an internal harness, not vendor benchmarks, with trajectory-level instrumentation that measures how agents succeed, not just pass@1.

The Production Data

Vercel's AI Gateway, spanning 200,000 teams over 7 months, reports 59% of all tokens are now agentic — multi-turn, tool-calling traces. This is production telemetry, not a forecast. The spend-volume split is where the routing behavior shows up: Anthropic captures 61% of dollar spend on reasoning and planning nodes via Opus, while Google captures 38% of token volume on high-throughput utility calls via Flash. Teams are already tiering by node type. Eval and cost code has not caught up.

Why Current Evals Are Measuring the Wrong Bottleneck

Most production eval harnesses still score single-turn responses against reference answers. When 59% of traffic is multi-step tool loops with retries, the failure mode is a planner burning 40,000 tokens arguing with itself, not final-answer accuracy. Accuracy is 90%+ in both cases. The bill lives on the cost path, which the harness does not see.

Cost models have the inverse problem. They were fit when input-output ratios sat around 3:1. Agentic traces run 15:1 on input with cache-hit rates that vary by provider. A forecast built on last year's ratio is off by roughly 5x on spend. That is not a calibration error. That is the wrong model.

If 59% of your tokens are agentic but 100% of your evals are single-turn, you're flying instruments-out.

The MDASH Validation

Microsoft's MDASH, a 100+ agent ensemble, outperformed Anthropic's Mythos on CyberGym by decomposing vulnerability discovery into scan → adversarial debate → PoC exploitation stages. The result is consistent with classical ML priors: ensemble topology with explicit disagreement beats monolithic models on complex tasks. Caveat: no cost or latency numbers published. The thing CyberGym doesn't measure is the inference bill for 100+ agents, which is the number that decides whether this ships.

The Architecture That Wins

The Capability Jump

UK AISI evaluated the newest Mythos and GPT-5.5-cyber variants on autonomous cyber-offense tasks. Mythos cleared both of AISI's hardest simulated attack ranges (full network takeover). GPT-5.5-cyber cleared one of two. The prior Mythos generation topped out at 'advanced persistence.' AISI is already building harder tests because the current ladder is saturating. The shape of the curve matters here. This is not smooth interpolation. It looks like a discrete unlock, comparable to GPT-3.5→4 on agentic tool-use.

The 271x Harness Delta

Same model, two teams, two orders of magnitude difference. Mozilla wrapped Claude Mythos Preview in a custom agentic harness integrated with their fuzzing infrastructure and surfaced 271 bugs in Firefox 150, including sandbox escapes, UAFs, and race conditions. Daniel Stenberg pointed the same model at curl and got 1 low-severity CVE with 4 false positives. His verdict: 'primarily marketing.'

The variable that moved was the harness, not the weights. Mozilla's wrapper emits reproducible test cases, scales across ephemeral VMs, and integrates into their security lifecycle. The thing the headline number doesn't tell you is which factor dominated: tool integration, compute budget per target, or corpus priors. On the evidence available, eval scaffolding dominates model choice by at least 50x on this task.

Vulnerability discovery just moved from human-weeks to model-minutes. If the patch SLA is benchmarked against human speed, it's tuned to last year's threat model.

Live Misuse Confirmed

Google's threat-intel team observed a hacking group using LLMs to build cybercrime tooling, the exact scenario flagged when Mythos shipped. Anthropic published a case study of Claude Code running an estimated 80-90% of tactical work across ~30 targets in what they call the first largely AI-executed espionage campaign. The 80-90% figure is self-reported and hard to audit, but the direction is consistent with Google's observation. This is no longer a red-team thought experiment.

What This Means for Release Gates

The Claims

The marginal dollar in model development looks like it is moving away from raw FLOP-hours and toward recipe design and data curation. The week's research drops point that direction unevenly, hitting different phases of the training pipeline.

Which One to Spike First

Token Superposition Training (TST) is the highest-leverage bet. It is a pretraining recipe change with no inference-side cost downstream. If it replicates at even 1.6x on a 1B continued-pretraining run with no val-loss regression, it pays for itself on the next full run. The 2-3x claim at matched FLOPs with no architecture change is either free or it is not, and the experiment is bounded.

Datology's result is the clearest evidence this year that data curation dominates compute for VLM training. Beating InternVL3.5-2B by about 10 points while using 17x less compute, purely through data selection, suggests most teams are over-spending on GPU-hours and under-investing in dataset engineering. Their 4B model lands near-frontier quality at 3.3x lower response FLOPs than Qwen3-VL-4B, which is a real serving-cost win, not a leaderboard one.

Star Elastic's 360x is the claim most likely to shrink under independent reproduction. Even at 30x it restructures how model-size tiers get produced for different deployment targets.

The marginal dollar in VLM training has moved from compute to curation. Datology's 17x result is the strongest evidence this year.

Adjacent: SWE-ZERO-12M-trajectories

Kevin Li released 112B tokens, 12M trajectories, 122K PRs, 3K repos, 16 languages, positioned as the largest open agentic trace corpus. The thing this doesn't tell you is durability: open releases at this scale tend to acquire licensing frictions within a few months. Useful for SFT, reward-model training, and synthetic eval generation while the license window is clean.