The Scientist

Three price moves landed simultaneously, and they compound

Anthropic ran a coordinated pricing reset this week. It hits any team running Claude at production scale in three ways:

1. Subscription-to-credit conversion: Claude paid plans now cap programmatic usage (Agent SDK, claude-p, GitHub Actions, third-party harnesses) at dollar-equivalent API credits. The implicit 70-90% discount on alternative-harness usage is gone.
2. June 15 third-party unbundling: Credits for Zed, Conductor, OpenCode, and T3 Code become a separate bucket equal to plan value. No rollover. Overflow bills at API rates.
3. Opus 4.7 image cost tripled: vision workloads that were unit-economic at the old price need re-evaluation against GPT-4V and Gemini before next sprint.

The capacity admission changes the story

Dario Amodei told Code with Claude that Anthropic planned for 10x growth and got 80x. An 8x forecast miss is the cleanest explanation for why Claude Code degraded through April, why rate limits tightened, and why Anthropic is now leasing xAI's entire Colossus 1 cluster (220,000+ GPUs) from a CEO who called them 'misanthropic and evil' three months ago.

ServiceNow's CDIO burned the full-year Claude budget by May. National Life Group's CIO called Claude 'great for consumer usage but not great for companies' that want per-user monitoring. The vendor provides no native per-user telemetry, no SLAs on latency or availability.

Any Claude benchmark from before May 7 is stale. The serving fleet is about to include heterogeneous hardware (H100, H200, GB200 via Colossus 1), which means p95/p99 latency variance will get weirder before it stabilizes.

OpenAI's counter-offensive

Sam Altman posted a 2-month-free Codex enterprise switch promo the same day Anthropic metered its users. Ramp data shows the enterprise race is genuinely contested at 34.4% vs 32.3%. The promotional window is asymmetric: free evaluation of the alternative with zero commitment cost. The thing the share number doesn't tell you is whether switchers stay after the two months expire. We will know by Q3.

The combined read for production stacks

The structural response is clear: gateway everything. No Claude call should leave infra without passing through a proxy that tags tenant_id, feature_id, user_id, and prompt-family hash. LiteLLM and Portkey get 80% of this in a day. The model vendor has explicitly offloaded observability to the customer.

The production traffic mix inverted without the tooling noticing

Vercel's AI Gateway production index is the first multi-tenant usage data worth anchoring on, and it reports 59% of all tokens are now agentic workloads. Six months ago the share was under 20%. This is not a forecast. It is current traffic across 200,000 production teams.

The spend-versus-volume split underneath the headline is where the architectural read sits:

The signature is tiered routing: expensive models plan, cheap models fan out. Teams paying Opus rates on every agent step are leaving 20-40% cost reduction unclaimed.

Why the eval harness is lying

Most harnesses still score single-turn responses against reference answers. Once 59% of tokens are multi-step tool loops with retries, the metrics that matter shift to:

• Tool-call precision/recall — did the agent select and invoke the right tools?
• Steps-to-completion — how many turns to resolve versus optimal?
• Cost-per-successful-task, not cost-per-token
• Recovery-from-error rate — does it self-correct or loop?

Final-answer accuracy lands at 90%+ in both monolithic and agentic architectures. The thing that number doesn't measure is the cost path. A planner that burns 40,000 tokens arguing with itself before succeeding is indistinguishable from a clean three-step resolution under pass/fail.

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.

Cost models have the same problem in reverse

Most cost forecasts were fit when input-output ratios sat near 3:1. Agentic traces run closer to 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. Glean reports off-the-shelf MCP burning 30% more tokens than retrieval-tuned knowledge graphs on comparable tasks. The methodology is undisclosed and the source is vendor-published, but the direction matches what verbose tool outputs do to context windows.

The routing architecture that matches production

Abridge's disclosure of 80M+ clinical conversations through a 'constellation of models' is the same pattern at healthcare scale: cheap fast model triages, larger model reasons only when needed. The 5-10x cost reduction at scale is the envelope. The open question is which steps degrade when downgraded, and only per-node quality measurement answers that.

Mythos cleared both AISI attack ranges. That's a first.

The UK AI Security Institute evaluated Anthropic's Claude Mythos and reports it is the first model to clear both simulated attack ranges, completing full network takeovers in controlled environments. The prior Mythos generation topped out at 'advanced persistence.' OpenAI's GPT-5.5-cyber cleared one of two ranges. AISI is already building harder tests, which is the tell that the current ladder is saturating.

This looks like a discrete capability unlock rather than smooth progress. The shape resembles GPT-3.5 to GPT-4 on tool-use benchmarks: not gradual interpolation, step-function behavior at a threshold. The thing this doesn't tell you is whether the threshold is the model or the scaffolding around it. AISI doesn't publish the harness, so we are reading a joint score.

The 271-vs-1 result is the methodology lesson

Mozilla wrapped a custom agentic harness around 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. Same weights. Two orders of magnitude difference in yield. The variable was the harness.

Microsoft's MDASH multi-model system shipped 16 real Windows fixes in May Patch Tuesday, which is the first count large enough to move LLM vulnerability discovery out of demo phase. DepthFirst claims 12 FFmpeg memory bugs for $1,000 against Mythos finding zero at $10K. Cost per bug is the relevant axis here, not leaderboard rank.

Meanwhile, the frameworks are being popped

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.

What this means for release gates

Gating agent releases on refusal rates or prompt-injection catch rates does not measure end-to-end chain completion, which is the bottleneck you actually hit in production. A staged rubric covering recon, initial access, lateral movement, persistence, and exfil, run against every model upgrade, is closer to the real failure mode. The AISI framing is a reasonable starting point even when sandboxed internally. Agents with shell access or repo write permissions need this eval before the next deployment.

Three drops that change the training cost calculus this quarter

Three independent research results landed in the same week, each attacking training efficiency from a different angle. The combined read: the marginal dollar in model development has moved from compute to recipe.

TST is the one to spike on first

Token Superposition Training is a pretraining recipe change with no inference-side downstream. If it replicates, it is a 2-3x wall-clock win for free. The mechanism is validated from 270M through 10B-A1B MoE scale. Because it leaves the inference architecture alone, it slots into an existing serving stack unchanged. The cheapest validation available is a 1B continued-pretraining spike against a matched-FLOPs baseline. The thing this doesn't tell you is how the gain behaves past 10B dense, which is where most teams actually care.

Datology: curation beats compute

The clearest evidence this year that the marginal dollar in VLM training has moved from compute to curation. Their 2B model beat InternVL3.5-2B by about 10 points at 17x less training compute, purely via data curation. The 4B variant lands near-frontier quality at 3.3x lower response FLOPs than Qwen3-VL-4B, which is a serving-cost win, not just a leaderboard one. Worth distinguishing: the training compute ratio is the headline, the inference FLOPs ratio is what shows up in your bill.

Star Elastic's 360x number is the kind of claim that always shrinks under independent eval. Even a 30x hold would restructure how you produce size tiers from a single post-training run.

Paired with the DuckDB and Kafka architecture shifts

On the infrastructure side, DuckDB's Quack HTTP protocol makes single-node analytics viable as a shared service. For sub-100GB workloads it is a credible replacement for Spark-on-Glue. Kafka Share Groups decouple consumer parallelism from partition count, with reported 8x throughput at 32 instances. The 8x figure is from the announcement, not a third party, so treat it as a ceiling until someone reproduces it. Both items matter here because they lower the infrastructure cost of the data pipeline feeding training runs.

The combined message: a lot of training compute is being wasted on orchestration overhead, suboptimal data, and infrastructure chosen when the ratio was different. Recipe wins compound with infrastructure wins.