AnthropicEndsClaudeSubscriptionDiscount,OpenAICounters

◆ DEEP DIVES

1. 01

   Anthropic's Triple Shock: Your Claude Cost Model Broke Overnight

   What Happened

   Anthropic shipped pricing changes inside a 48-hour window that compound on each other. First, all programmatic subscription usage (Agent SDK, claude-p, GitHub Actions, third-party harnesses) converted from flat-rate to dollar-matched API credits. The 70-90% effective discount power users had been quietly running on is gone. Second, starting June 15, third-party tool usage (Zed, Conductor, OpenCode, T3 Code) lands in a separate credit bucket with no rollover. Overflow bills at API list rates. Third, Opus 4.7 tripled image processing costs.

   The capacity numbers explain the timing. Dario Amodei conceded Anthropic planned for 10x growth and got 80x in revenue and usage. The gap forced an emergency lease of xAI's entire Colossus 1 cluster — 220,000+ GPUs across H100, H200, and GB200. That is roughly 45% of xAI's current capacity changing hands.

   ---

   Why It Matters Now

   ServiceNow's CDIO burned through the full-year Claude budget by May. National Life Group's CIO publicly called Claude 'not great for companies' wanting per-user monitoring. Both complaints route to the same root cause. Anthropic provides no native per-user, per-tool usage telemetry and no SLAs on latency or availability. Customers wire their own observability or fly blind.

   If the vendor cannot tell you which user burned the token, the problem is not cost. It is observability, and it is yours to fix before the next invoice.

   OpenAI's response shipped the same day: a 2-month-free Codex enterprise switch promotion aimed at developers Anthropic just repriced. Ramp puts Anthropic at 34.4% of paying businesses against OpenAI at 32.3%, the first crossover. The thing this doesn't tell you is how the share moves once the new pricing hits actual invoices, which is the measurement that matters. The market is contested. Switching costs are the only moat left.

   ---

   The Capacity Math

   Surface	Before	After (May 7-14)
   Claude Code limits	5-hour cap	Doubled
   Peak-hours throttle	Reduced limits	Removed
   Opus API rate limits	Squeezed	'Substantially raised'
   Fleet composition	Anthropic-managed	+220K GPUs via Colossus

   Any Claude benchmark from before May 7 is stale. Architectural decisions made on April or May numbers describe a system that no longer exists; aggressive caching keyed to the old discount structure is the obvious example. Re-baseline after the new caps land.

   Action items

   • Audit every Claude-backed workload (Agent SDK, GitHub Actions, batch evals) and reconcile projected token burn against the new credit cap before month-end
   • Deploy an LLM gateway (LiteLLM/Portkey) with per-user, per-feature tagging and daily token budget alerts this sprint
   • Run a 2-month Codex evaluation under OpenAI's free enterprise switch promo using matched prompts and your own eval harness
   • Avoid long-term Anthropic commits (annual contracts, dedicated capacity) until post-Colossus integration stability is observable in 6-8 weeks

   Sources:Claude just metered your agent SDK calls · Claude Code latency on long-context requests · Anthropic ships no per-user usage telemetry · Anthropic passes OpenAI in B2B · Vercel published a number worth sitting with · Ramp's AI Index shows Anthropic at 34.4%

2. 02

   59% Agentic Traffic — Your Eval Harness Is Scoring the Minority

   The Production Reality

   Vercel's AI Gateway production index, across 200,000 teams over 7 months, reports 59% of all tokens are now agentic — multi-turn, tool-calling traces. Six months ago that figure was under 20%. Most cost models in use 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 built on last year's ratio is off by roughly 5x on spend.

   The eval harness mismatch is worse. Most harnesses still score single-turn responses against reference answers. The thing this doesn't tell you is what happens when the median request is a multi-step tool loop with retries, and the failure mode is a planner burning 40,000 tokens arguing with itself before giving up. Final-answer accuracy reads 90%+ in both the healthy and the pathological case. The bill, not the accuracy, is where the failure lives.

   ---

   The Spend-Volume Divergence

   Provider	Share of Spend	Share of Volume	Primary Model	Implied Role
   Anthropic	61%	—	Opus	Reasoning / planning nodes
   Google	—	38%	Flash	High-throughput utility calls
   Others	~39%	~62%	Mixed	Mixed

   That is a textbook tiered-routing signature. Expensive models handle planning and reasoning; cheap models do retrieval rewriting, extraction, and classification. Running Opus on every agent step leaves 20-40% of inference spend on the table at constant trajectory completion rate. The condition on that estimate is a stable router policy and unchanged tool surface; rewrite either and you re-measure.

   ---

   The Duolingo Anchor

   Duolingo's CEO pegged AI-generated content 'slop' at ~20% requiring human QC. That is a rare production quality number from a deployment at volume, and it is the anchor an internal harness should be calibrated against. The same company reversed its blanket 'evaluate all employees on AI usage' policy after observing performative adoption without productivity lift. Goodhart's Law arrived on schedule.

   LLM-as-Verifier vs. LLM-as-Judge

   A methodology paper this week argues that decomposing evaluation into repeated binary verifications at token granularity eliminates the tie-rate problem in pairwise judges. The mechanism is mathematically clean: one high-variance categorical judgment replaced by k lower-variance Bernoulli tests. Worth a one-day swap on one eval pipeline to measure the tie-rate reduction before committing further.

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

   Action items

   • Add trajectory-level metrics (tool-call F1, steps-to-completion, cost-per-successful-task) to your eval harness alongside single-turn benchmarks this sprint
   • Instrument per-node token cost in agent graphs and route utility calls (JSON extraction, query rewriting) to Flash/Haiku-class models
   • Benchmark your LLM output acceptance rate against Duolingo's ~20% slop disclosure; adjust HITL budget if delta exceeds 10pp in either direction
   • Prototype LLM-as-Verifier on one existing eval pipeline; measure tie-rate and inter-run variance against current judge setup

   Sources:Agentic traffic crossed fifty-nine percent · Vercel published a number worth sitting with · The CyberGym result · Duolingo's twenty percent AI slop rate · LLM-as-a-Verifier reframes evaluation

3. 03

   Training Efficiency Moved 2-360x — The Unit Economics of Your Next Run Just Changed

   Three Results, Same Direction

   Training unit economics shifted on three fronts this week, all pointing the same way. Each result matters for any team running its own pretraining, fine-tuning, or distillation.

   Work	Claim	Scale Validated	Inference Impact	Replication Risk
   Nous TST	2-3x wall-clock at matched FLOPs	270M → 10B-A1B MoE	None — no arch change	Medium; single-source
   NVIDIA Star Elastic	360x cheaper model-family derivation	Not specified	Produces family from one run	High; lab-reported
   Datology VLM	+11.7 pts, 17x less compute	2B and 4B params	Lower response FLOPs	Medium; benchmark-selection risk

   ---

   Which One to Spike First

   Token Superposition Training (TST) is the cleanest opportunity. It is a pretraining recipe change with no inference-side downstream. If it replicates, it is a free 2-3x on wall-clock with no new serving complexity, since the architecture at serve time is unchanged. The validation range, 270M to 10B, covers most continued-pretraining workloads teams actually run.

   Star Elastic's 360x is the kind of headline number that always shrinks under independent eval. Given the gap between paper conditions and production data, I would plan around a 30x hold and treat anything above that as upside. Even at 30x, one post-training run producing the full edge/mobile/server family eliminates the multi-run coordination overhead teams currently eat to maintain deployment tiers.

   Datology's result is the clearest evidence this year that the marginal dollar in VLM training has moved from compute to curation. They beat InternVL3.5-2B by ~10 points at 17x less training compute, purely via data selection. The thing this doesn't tell you on its own is the serving side: the near-frontier 4B model achieves 3.3x lower response FLOPs than Qwen3-VL-4B, which is a real inference-cost win on top of the training-cost win.

   ---

   The DuckDB + Kafka Infrastructure Shift

   Two data-infrastructure releases land on the same architectural assumption. DuckDB's Quack HTTP protocol turns embedded analytics into a proper shared service, and is a credible replacement for Spark-on-Glue for single-node ETL under ~100GB. Kafka Share Groups break the partition==max-consumers ceiling with ~linear 8x scaling at 32 instances on I/O-bound workloads. Combined with the ECS Fargate + Terraform pattern, that is a template for deleting the distributed-compute costume a lot of single-node jobs are wearing.

   VLM training value moved from compute to curation. Data pipelines are following a parallel move from cluster orchestration to statistics hygiene.

   Action items

   • Spike Token Superposition Training on a 1B continued-pretraining run against a matched-FLOPs baseline this quarter
   • Audit Glue/EMR job catalog for single-node candidates (<100GB working set) and port one onto ECS Fargate + DuckDB + Terraform
   • Run ANALYZE/compute-stats coverage audit across Iceberg/Delta tables; add stats maintenance to table-level SLAs
   • Benchmark Kafka Share Groups against most partition-bound consumer group (embedding/enrichment workloads first)

   Sources:DuckDB shipped a client-server mode this week · Claude just metered your agent SDK calls · DuckDB + Quack plus ECS Fargate/DuckDB/Terraform ETL pattern

4. 04

   AI Cyber Offense Hit Full Network Takeover — Your Eval Harness Doesn't Measure This

   The Capability Threshold

   UK AISI confirmed that Anthropic's Mythos cleared both simulated attack ranges, the first model to do so. The prior generation topped out at 'advanced persistence.' GPT-5.5-cyber cleared one of two. AISI is already building harder tests because the current ones are saturating, which is what you'd expect when the leaderboard catches up to the eval rather than the underlying capability. Separately, Google's threat-intel team observed a hacking group using AI to build cybercrime tooling in the wild. That is the first production-grade confirmation of agentic misuse moving from red-team hypothesis to detected incident.

   ---

   The Harness Delta That Matters

   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 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 that moved was the harness.

   Dimension	Mozilla + Firefox	Stenberg + curl
   Model	Mythos Preview	Mythos Preview
   Harness	Custom agentic, fuzzer-integrated	Out-of-box scan
   Bugs surfaced	271	5 claimed (4 FP)
   True positives	Acted upon	1 low-severity CVE

   The implication for anyone shipping agents: eval budget dominates model choice by 50x or more. The thing the headline benchmark doesn't tell you is which harness produced it. A week of domain-specific scorer and reproducible test-case emitter work looks like a better trade than the next dollar of inference.

   ---

   New CVEs Targeting Your Stack This Week

   Beyond the previously reported LiteLLM KEV, new critical vulnerabilities landed on infrastructure data teams actually run:

   • Apache Iceberg (CVE-2026-42812, CVSS 9.9): attackers redirect table metadata writes to attacker-chosen S3 prefixes. Next query reads poisoned Parquet. Next training run ingests corrupted features.
   • Apache Polaris (CVE-2026-42809/10/11, CVSS 9.9): credential-broadening enables cross-tenant cloud access.
   • Argo CD (3.2.x/3.3.x, CVSS 9.6): read-only users extract plaintext K8s Secrets including model-registry tokens and cloud credentials.
   • NGINX rewrite module: 18-year latent unauthenticated RCE affecting every model-serving gateway behind NGINX.

   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.

   Action items

   • Patch Argo CD to ≥3.2.12 / ≥3.3.10 and rotate every K8s Secret in reachable namespaces this week
   • Audit Iceberg/Polaris catalog configurations: enforce write-path allowlisting for table metadata locations and rotate exposed credentials
   • Patch NGINX across all inference gateways; audit rewrite-module usage in model routing configs
   • Add autonomous-cyber-capability tier to model eval harness (AISI-style attack-range tasks) before next model upgrade

   Sources:LiteLLM landed in the KEV catalog · Mythos cleared the AISI attack ranges · PraisonAI exploited in 4 hours · Mozilla shipped 271 bugs · Google's report of a threat actor using AI · The UK AISI evaluations report full network takeovers