The Scientist

What Happened

The headline change: Claude subscriptions now convert to dollar-matched API credits for all programmatic usage, including Agent SDK, claude-p, GitHub Actions, and third-party harnesses. The implicit 70-90% effective discount is gone. In the same week, Opus 4.7 tripled image processing cost, and starting June 15, third-party tool usage (Zed, Conductor, OpenCode, T3 Code) moves to a separate credit bucket equal to plan value with no rollover and overflow at API rates.

The driver behind the pricing is capacity. Anthropic planned for 10x growth and is seeing 80x. The emergency fix is leasing xAI's entire Colossus 1 cluster, 220,000+ GPUs spanning H100, H200, and GB200. A CFO is in seat and the company is targeting an October IPO, which is a reasonable proxy for why margin-per-token is now a board metric.

The Capacity Context

The pricing changes read more cleanly alongside the capacity numbers. ServiceNow's CDIO burned through a full-year Claude budget by May. National Life Group's CIO called Claude 'not great for companies that want per-user monitoring,' and Anthropic ships no native per-user telemetry and no SLAs, which is unusual for a dependency sitting on production critical paths.

Any Claude benchmark from before May 7 is stale, and any cost model built against flat subscription rates is not directionally wrong, it is numerically wrong.

The OpenAI Counter-Move

Sam Altman posted a 2-month-free Codex enterprise switch promo on the same day Anthropic metered subscriptions. Ramp's April data showed Anthropic edging OpenAI for the first time, 34.4% vs 32.3%. The promo is an asymmetric-payoff bet: free to evaluate, with bounded switching cost if you already have a provider abstraction layer. OpenAI is pricing directly into the developer cohort Anthropic just alienated.

Cross-Source Tension

Sources disagree on the durability of Anthropic's lead. Ramp data is a card-spend proxy and measures who gets billed, not token volume or production criticality. A 20-seat pilot weighs the same as a company at inference scale, and OpenAI notes large enterprises rarely pay by card. The directional signal looks real; the magnitude does not. What is not uncertain is that the market is now genuinely multi-vendor, and architecture should reflect that.

The Production Shift, Quantified

Vercel's AI Gateway telemetry — 200,000 teams over 7 months — reports that 59% of all token volume is now agentic: multi-turn, tool-calling traces, not single-shot completions. Six months ago the share was under 20%. The interesting structure is the spend-volume split. Anthropic captures 61% of spend via Opus on reasoning and planning nodes, while Google captures 38% of volume via Flash on throughput and utility calls. Vendor loyalty is not visible in the data. Customers churn freely.

This is not a leaderboard result. It is production telemetry from a multi-tenant gateway, which means most tokens in the wild are inside multi-step tool loops with retries, not the single-turn completions an eval harness was built to score.

What Breaks at 59%

Eval harnesses: single-turn accuracy on held-out prompts measures the 41% minority. The thing this doesn't tell you is whether a planner burns 40K tokens arguing with itself before giving up. Pass rate looks fine. Spend is 5x budget. Cost models: most were calibrated when input-output ratios sat at 3:1. Agentic traces run at ~15:1 on input, with heavy cache reuse on some providers and none on others. A forecast carried over from last year's ratio is off by roughly 5x on spend.

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.

The Routing Architecture That Fits

The Vercel pattern is consistent with what Abridge disclosed across 80M+ clinical conversations: a constellation of models, cheap fast triage in front, expensive reasoning behind, per-task selection. Given those numbers, the plausible envelope for cost reduction from tiered routing is 20-40% at constant trajectory completion rate. Glean's published benchmark puts off-the-shelf MCP at 30% more tokens than retrieval-tuned knowledge graphs. The methodology is not disclosed and the source is the vendor, so treat it as directional. It is consistent with the MCP context-window bloat people have measured independently.

The Enterprise Convergence

SAP (€100M partner investment) and ServiceNow (Action Fabric) have landed in the same place: agents need Knowledge Graph grounding + MCP-exposed workflows. The architectural drift is from RAG-over-docs to structured, entity-resolved context. The eval needs to move with it. Tool-use accuracy against grounded context is the production differentiator. It is on no public leaderboard.

Three Independent Wins, One Direction

The marginal dollar in model development has moved from raw compute to architecture and curation. The evidence this week comes from three separate labs, each working a different layer of the training stack, and the results point the same way. The marginal dollar in model development has moved from raw compute to architecture and curation.

TST Is the One to Spike First

Token Superposition Training is a pretraining recipe change with no inference-side downstream. If it replicates, it is a 2-3x improvement on wall-clock for free. The serving architecture is unchanged. The efficiency lives entirely in how gradients are computed during training. Validated from 270M up to a 10B-A1B MoE, which is the band most teams actually run for continued pretraining and domain adaptation.

Datology: Curation Beats Compute

Datology's result is the cleanest evidence so far that data curation now dominates compute scaling for VLMs. A 2B model lands about 10 points above InternVL3.5-2B at 17x less training compute, on data selection alone. At 4B, they reach near-frontier quality at 3.3x lower response FLOPs than Qwen3-VL-4B, which is a serving win and not just a training one. For any team maintaining a vision pipeline, this reorders the next budget line. Curation tooling before more GPUs.

Star Elastic: Speculative but High-Leverage

NVIDIA's claim that one post-training run produces a full family of model sizes at 360x lower cost than pretraining each is the kind of number that always shrinks under independent evaluation. The thing this number doesn't tell you is what holds at the small end versus the large end. Even a 30x hold would restructure how teams produce size tiers for routing. Paired with the 59% agentic routing pattern, cheap access to a 1B-to-70B spectrum from a single run is what a tiered architecture actually needs.

TST is worth a spike on your next continued-pretraining run. If wall-clock comes in at even 1.6x with no val-loss regression, it pays for itself immediately.

The Parallel Signal: Only 15% Are Ready

Fivetran's readiness index reports that 15% of organizations have the data foundation for agentic AI. Data quality and lineage are the top blocker, cited by roughly 50%. Read against the efficiency results, the picture is consistent. Training is getting cheaper. Most teams cannot capitalize because the data layer is not there. Half the agent projects funded this quarter are data-platform projects with an agent bolted on.