The Scientist

What Changed This Week

Anthropic shipped three changes simultaneously that compound into one budgeting problem. First, all Claude subscriptions now convert to dollar-matched API credits across Agent SDK, claude-p, GitHub Actions, and third-party harnesses. The implicit 70-90% subsidy that power users extracted from Max plans is gone. Second, on June 15, Claude usage through third-party tools (Conductor, Zed, OpenCode, T3 Code) moves to a separate credit bucket with no rollover. Overflow bills at list API rates. Third, Dario Amodei conceded Anthropic planned for 10x growth and got 80x, which explains months of quiet degradation and the emergency lease of xAI's full 220,000-GPU Colossus 1 cluster.

Why This Breaks Your Stack

The metering change is not a gentle price increase. It is a structural reclassification of how programmatic usage is billed. Any batch eval, enrichment pipeline, or agent loop running on a flat subscription is now burning metered tokens at list price. ServiceNow's CDIO confirmed publicly that they burned their full-year Claude budget by May, after price hikes hit an enterprise with no native per-user telemetry.

Anthropic provides no native per-user, per-tool usage attribution. Customers must wire external analytics to see who is consuming what.

The capacity story compounds the pricing story. The 80x miss means serving conditions your eval harness measured between mid-summer and now are contaminated for baselining. Rate limits on Opus are being raised, Claude Code 5-hour caps are doubling, and a heterogeneous fleet (H100 + H200 + GB200 via Colossus) means p95/p99 latency variance will increase during integration, not decrease. The thing this doesn't tell you is which slice of your traffic lands on which silicon.

The OpenAI Counter-Move

Sam Altman posted a 2-month-free Codex enterprise switch promo the same day Anthropic announced metering. Ramp's April data shows Anthropic edging OpenAI 34.4% vs 32.3%, the first apparent lead change. OpenAI is pricing explicitly against the developers Anthropic just alienated. The asymmetric free evaluation window expires in roughly 60 days.

What the Market Data Actually Shows

Ramp measures who gets billed, not token volume or production criticality. A 210-basis-point gap in a monthly snapshot is inside noise. The directional signal, that second-vendor adoption is now the default, is the actionable read.

The Production Data

Vercel's AI Gateway, covering 200,000 teams over 7 months, reports that 59% of all tokens are now agentic — multi-turn, tool-calling traces rather than single-shot completions. Six months ago the figure was under 20%. The spend/volume split is the more interesting cut: Anthropic captures 61% of spend via Opus on expensive reasoning, while Google captures 38% of volume via Flash on cheap fan-out. The data shows no vendor loyalty. Customers route by task.

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

Why Single-Turn Evals Are Now Measuring the Minority

Agentic workloads are not 'just more tokens.' They are bursty, multi-turn, tool-calling, with 15:1 input-to-output ratios against the 3:1 most cost models assume. The cost function is latency-per-step times number-of-steps, plus retry risk from bad tool calls. A forecast built on last year's ratio is off by roughly 5x on spend, and the error is not symmetric across vendors. The median request stops being the right summary statistic. p95 does the work.

MDASH Validates Multi-Agent Decomposition

Microsoft's MDASH, a 100+ agent system, beat Anthropic's Mythos on the CyberGym vulnerability benchmark by decomposing into scan → adversarial debate → PoC construction stages. The thing this doesn't tell you is which stage drives the lift, or what it cost. There is no ablation and no cost comparison. The architectural pattern is consistent with ensemble priors — specialized agents with explicit disagreement tend to generalize better than monolithic calls — but consistent-with is not evidence-for.

The Abridge Reference Architecture

Abridge runs 80M+ clinical conversations through a 'constellation of models': cheap triage in front, expensive reasoning behind, LLM judges calibrated against human annotators, memory externalized to event-driven stores rather than model weights. The transferable production pattern is the confidence-gated router. Only 40% of requests reach frontier-class reasoning. The rest are handled by 7-13B models at 5-10x lower cost.

What the 30% MCP Overhead Means

Glean's benchmark is vendor-sponsored with undisclosed methodology. It claims off-the-shelf MCP uses 30% more tokens and loses 2.5x head-to-head against a tuned knowledge graph on agentic tasks. Treat that as a hypothesis, not a result. The failure mode is plausible regardless: MCP tool listings inflate context windows, and naive tool outputs return verbose blobs where a reranked snippet would do. The 30% number is one to falsify on your own workload this sprint, not to cite.

The Landscape

Three research drops landed in the same cycle, each targeting a different cost bottleneck in the training pipeline. Individually, each is worth a spike. Together they suggest the marginal dollar in training has moved from raw compute to recipe optimization and data curation.

Nous Research: Token Superposition Training (TST)

TST reports 2-3x wall-clock speedup at matched FLOPs with no inference-time architecture change, validated from 270M through 10B-A1B MoE scale. The mechanism superimposes multiple token sequences during training, lifting effective throughput without touching the served model structure.

Why spike this first: nothing changes downstream at serving. If it replicates on a 1B continued-pretraining run, it is a free 2-3x on every subsequent training run. Replication risk is medium. Single source, but the claim is clean and falsifiable.

Datology: Data Curation Beats Compute

At 2B params, Datology reports +11.7 points on 20 VLM benchmarks, beating InternVL3.5-2B by about 10 points at 17x less training compute, purely through data curation. At 4B params, they hit near-frontier quality at 3.3x lower response FLOPs than Qwen3-VL-4B.

This is the clearest evidence this year that the marginal dollar in VLM training has moved from compute to curation.

The serving win is real. Lower response FLOPs means cheaper inference on every request. The risk is benchmark-selection bias. Twenty benchmarks sounds broad, but the specific selection matters, and the thing this doesn't tell you is how the curated data performs on slices outside that suite.

NVIDIA Star Elastic

Claims one post-training run produces a family of reasoning model sizes at 360x lower cost than pretraining a family, and 7x better than SOTA compression. Scale not specified in the ledger.

Caveat: this is the kind of headline number that shrinks under independent eval. Even a 30x hold would restructure how teams produce size tiers for deployment. 360x from a lab-reported result deserves heavy discounting until reproduced.

Comparative Assessment

What This Means for GPU Budget Planning

Against the 4:1 demand-to-supply ratio at Nebius and capacity selling out quarterly, these results offer a hedge: recipe-level efficiency gains can partially offset a capacity squeeze. TST alone, if it replicates at 2x, means the same training budget buys twice the runs. Combined with Datology's curation thesis, the highest-leverage Q3 investment is likely a data quality team, not more GPU hours.

The 270x Yield Gap

Two teams ran Claude Mythos Preview against large C codebases in the same window. Mozilla wrapped a custom agentic harness around their existing fuzzing infrastructure and reported 271 bugs in Firefox 150, including sandbox escapes, use-after-frees, and race conditions. Daniel Stenberg pointed the same model at curl with a generic scanner and got exactly 1 low-severity CVE alongside 4 false positives.

Same weights. Roughly two orders of magnitude in yield. The variable that moved was the scaffolding.

When a frontier model yields 271 bugs for one team and 1 CVE for another against the same language, the harness is the product, not the model.

Why This Generalizes Beyond Security

The pattern maps onto ordinary ML evaluation. A team debating Claude 4 vs. GPT-5 vs. Gemini for an internal tool is optimizing the wrong variable. On this evidence, the gap puts harness investment ahead of model selection by at least 50x. Mozilla's harness emits reproducible test cases, scales across ephemeral VMs, and feeds existing signal pipelines. The model contributes capability. The harness converts capability into measurable signal. The thing the 271 number doesn't tell you is how much of that yield is unique vs. duplicate clusters, which would tighten the multiple. It probably loosens it instead.

The AISI Threshold Crossing

Mythos is also the first model to clear both UK AISI simulated attack ranges, meaning full network takeover rather than advanced persistence alone. GPT-5.5-cyber cleared one of two. AISI is building harder tests because the current ladder is saturating, which is consistent with a discrete capability unlock rather than smooth interpolation. The closest analog is GPT-3.5 to GPT-4 on agentic benchmarks. It is possible the gap is narrower than two ranges suggests once variance is accounted for, but two-of-two vs. one-of-two is hard to explain by noise.

Operational implications for agent deployments

LLM-as-Verifier: A Methodological Upgrade

A parallel finding reinforces the harness thesis. LLM-as-a-Verifier beats LLM-as-a-Judge on tie-rate and decision accuracy by decomposing evaluation into repeated binary verifications with token-level scoring. The mechanism is straightforward: one high-variance categorical judgment replaced by k lower-variance binary ones. The same argument moved human eval from Likert scales to pairwise preferences a decade ago.

Practical translation: rewrite one pairwise judge as a decomposed verifier, measure tie rate before and after, and the width of the bootstrap CI on a known A/B pair. If CIs tighten at equal compute, roll to the rest of the suite. Cheapest variance reduction available this quarter.