AnthropicEndsClaudeSubscriptionDiscount,OpenAICounters

◆ DEEP DIVES

1. 01

   Anthropic's Triple Squeeze: Your Claude Budget Just Broke

   What Happened

   The pricing change is the surface event, and it compounds with two others. Anthropic metered all programmatic Claude usage: subscriptions now convert to dollar-matched API credits across Agent SDK, claude-p, GitHub Actions, and third-party harnesses. The implicit 70-90% discount power users were getting is gone. At Code with Claude, Dario Amodei disclosed that Anthropic planned for 10x growth and got 80x, which is the most plausible read on the quality degradation users have been logging since mid-April. The emergency fix is leasing xAI's entire Colossus 1 cluster (220,000+ GPUs), from the CEO who called Anthropic 'misanthropic' three months ago.

   The June 15 Cliff

   Starting June 15, Claude usage through third-party tools (Conductor, Zed, OpenCode, T3 Code) draws from a separate credit bucket equal to plan value. No subsidized tokens, no rollover, overflow billed at API rates. Claude Code and the Claude app are unaffected. Weekly rate limits get bumped fifty percent for two months as a softener.

   Surface	Before	After (announced)
   Claude Code limits	5-hour cap	Doubled
   Peak-hours throttle	Reduced for Pro/Max	Removed
   Opus API rate limits	Squeezed	'Substantially raised'
   Third-party tools	Subsidized	Credit-capped, overflow at list

   The Counter-Offensive

   OpenAI dropped a 2-month-free Codex enterprise switch promo the same day. Ramp's April data shows Anthropic edging OpenAI 34.4% vs 32.3% in B2B spend, the first apparent lead change. The thing the Ramp number doesn't tell you is whether the lead survives the metering change. OpenAI is pricing a response against the exact developers Anthropic just alienated, and a free evaluation window is an asymmetric-payoff bet for the buyer.

   Why This Matters For Your Stack

   ServiceNow already burned its full-year Claude budget by May after the price hikes. Anthropic provides no native per-user or per-tool usage telemetry and no SLAs, which is unusual for a dependency on the critical path. The vendor dashboard cannot attribute spend to a tenant or a prompt. That instrumentation has to come from the consumer side.

   Any Claude benchmark from before May 7 is stale. Any cost model built before the metering change is numerically wrong, not directionally wrong.

   The Capacity Aftermath

   The 80x growth miss means observed 'quality degradation' was quantization, smaller-model routing under load, and scheduler unfairness, not product changes. The Colossus integration (H100, H200, GB200 heterogeneous fleet) will produce p95/p99 variance during transition. Tail latency gets weirder before it stabilizes, and that is the metric to watch, not the mean.

   Action items

   • Audit every Claude-backed workload (Agent SDK, GitHub Actions, batch evals) against the new credit cap and flag jobs that exhaust credits before month-end
   • Deploy an LLM gateway (LiteLLM/Portkey) with per-user, per-feature tagging and daily token budget alerts in front of all Claude traffic
   • Run a 2-month Codex evaluation under OpenAI's free promo with matched prompts and tool schemas against your existing Claude harness
   • Re-baseline Claude throughput and latency benchmarks after Colossus integration stabilizes (target: late May)
   • Reforecast Claude spend for teams using Zed, Conductor, or OpenCode — model post-June-15 scenario where overflow hits API rates

   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 · Anthropic passed OpenAI in enterprise share

2. 02

   59% Agentic Tokens: Your Eval Harness and Cost Model Are Measuring the Minority

   The Production Data

   Vercel's AI Gateway, covering 200,000 teams over 7 months, reports 59% of production tokens as agentic now: multi-turn, tool-calling traces rather than single-shot completions. Six months ago the figure was under 20%. The composition has shifted faster than at any point since chat endpoints replaced completion endpoints.

   The Spend-Volume Divergence

   The provider landscape has split along two axes that do not line up:

   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

   Teams are already routing across providers: expensive models for reasoning, cheap models for throughput. The Gateway data shows no vendor loyalty. A serving layer hardcoded to one provider SDK is out of step with what 200K production teams are actually running.

   Why Existing Harnesses Fail

   Most agent eval harnesses score single-turn responses against reference answers. That was the right call in 2023. Agentic traces now run at a 15:1 input-to-output ratio, with heavy cache reuse on some providers and none on others. A forecast built on last year's 3:1 ratio is off by roughly 5x on spend. The metrics that matter now:

   • Cost-per-successful-task, not cost-per-token
   • Tool-call precision and recall, not single-turn accuracy
   • Steps-to-completion and retry rates
   • Recovery-from-error rate across the trajectory

   Glean's benchmark claims off-the-shelf MCP burns 30% more tokens than retrieval-tuned knowledge graphs on agentic tasks. Vendor-published with no methodology, so treat it as a hypothesis. The thing this doesn't tell you is whether the gap holds outside Glean's own task distribution. The failure mode is real, though: MCP tool listings balloon context windows, and naive tool outputs return verbose blobs where a reranked snippet would do.

   The Routing Architecture That Wins

   Abridge (80M+ clinical conversations, 250 health systems) discloses the pattern: cheap fast model for triage, large model for reasoning only when needed. Post-training on proprietary data beats frontier on narrow tasks at lower cost. The plausible cost reduction envelope from tiered routing is 20-40% at constant trajectory completion rate. If it lands at half that, the routing layer still pays for itself.

   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 (task success, tool-call F1, steps-to-completion, cost-per-successful-task) to your eval harness this sprint
   • Instrument per-node token cost in your agent graph and route utility calls (summarization, JSON extraction, query rewriting) to Flash/Haiku-class models
   • Run a 1-hour spike measuring token overhead of your current MCP/tool-calling setup vs. a retrieval-first baseline on 100 production traces
   • Add LLM-judge-to-human-annotator agreement as a tracked quarterly metric in your eval pipeline

   Sources:Agentic traffic crossed fifty-nine percent · Vercel published a number worth sitting with · The CyberGym result · Abridge runs model routing across 100M conversations · MCP plus knowledge graphs

3. 03

   Three Training Efficiency Results That Change Your Q3 Compute Budget

   The Drops

   A cluster of pretraining results landed in one week, each nudging training unit economics in a direction that matters for any team running pretraining, fine-tuning, or distillation.

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

   Why TST Is the One to Spike First

   Token Superposition Training is a pretraining recipe change with no inference-side downstream. If it replicates, it's a free 2-3x speedup. The mechanism, superposing multiple token targets per forward pass, has been validated from 270M to 10B-A1B MoE scale. No architecture change means existing serving infrastructure stays untouched. This is the cleanest 'try it and know within a week' experiment on the list.

   Datology: The Marginal Dollar Moved to Curation

   Datology's result is 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 ~10 points across 20 benchmarks while using 17x less training compute, and their 4B matched near-frontier quality at 3.3x lower response FLOPs than Qwen3-VL-4B. The serving cost reduction is real. The thing this doesn't tell you is whether the curation advantage is durable or copied within two quarters.

   Star Elastic's 360x Needs Discount

   NVIDIA's claim that one post-training run produces a family of reasoning model sizes at 360x lower cost than pretraining a family is the kind of headline number that always shrinks under independent evaluation. Even a 30x hold would restructure how size tiers get produced. But 360x→30x is a common compression pattern in lab-reported claims. Wait for replication before planning around it.

   Combined Implication

   Together these results suggest that Q3 2026 training runs priced at Q1 2026 efficiency assumptions are systematically overspending. A team about to allocate $X to a continued-pretraining run should first check whether TST halves the wall-clock, whether curation halves the data requirement, and whether Star Elastic removes the need to pretrain multiple sizes. The order matters: TST replicates fastest, Datology requires a curation pipeline you may not have, and Star Elastic is still on the wait-for-replication shelf.

   The marginal dollar in VLM training moved from compute to curation. The marginal dollar in pretraining moved from more FLOPs to better recipes. Budgets that don't reflect this are overpaying.

   Action items

   • Spike Token Superposition Training on a 1B continued-pretraining run against a matched-FLOPs baseline this month
   • Audit your VLM/multimodal training data pipeline for curation quality — measure per-shard quality scores and deduplicate before the next training run
   • Hold off on pretraining separate model size tiers until Star Elastic publishes reproducible details
   • Pull SWE-ZERO-12M-trajectories (112B tokens, 12M trajectories, 3K repos) and begin preprocessing before licensing frictions appear

   Sources:Claude just metered your agent SDK calls · DuckDB shipped a client-server mode this week

4. 04

   Apache Iceberg/Polaris CVSS 9.9 and Argo CD 9.6: The Lakehouse Trust Boundary Shrank

   What's New This Week

   On top of the LiteLLM (KEV-listed) and Ollama exploits flagged last week, four more critical CVEs landed on the data and ML stack in the same window. The blast radius is the infrastructure most DS teams treat as trusted by default, which is the part that matters.

   The New CVEs

   Component	CVE / CVSS	Attack Class	Blast Radius
   Apache Iceberg	CVE-2026-42812 / 9.9	Metadata write redirect	Poisoned tables, corrupted training data
   Apache Polaris	CVE-2026-42809/10/11 / 9.9	Credential broadening	S3/GCS creds, cross-tenant access
   Argo CD 3.2.x/3.3.x	CVE-2026-42880 / 9.6	Missing authorization	Plaintext K8s Secret extraction
   PraisonAI	CVE-2026-44338	Auth bypass	Agent runtime, connected DBs
   NGINX rewrite	18-year latent RCE	Unauthenticated RCE	Model-serving ingress

   Why Iceberg CVSS 9.9 Is the Worst

   An attacker with table-write permission can point metadata at an attacker-controlled S3 prefix. The next query reads poisoned Parquet. The next training run ingests silently corrupted features. The thing CVSS doesn't tell you is that most lakehouse observability does not detect metadata pointer mutations. Default logging covers row changes, not pointer changes. Combined with Polaris credential-broadening, there is a plausible path from compromised analyst notebook to cross-tenant data theft.

   Argo CD: Every Secret Is Disclosed

   For teams running training jobs or model services on Argo CD 3.2 or 3.3, treat every K8s Secret in reachable namespaces as disclosed until patched and rotated. That includes model-registry tokens, HuggingFace PATs, DB passwords, and cloud credentials. The rotation costs more than the patch. It is also the part that actually closes the incident.

   PraisonAI: Agent Frameworks Are First-Class Targets

   PraisonAI's auth bypass was weaponized within 4 hours of CVE disclosure. Four hours is the operational signal: agent frameworks have crossed the adoption threshold where threat actors monitor their disclosure feeds. Any team prototyping with LangChain, CrewAI, AutoGen, or PraisonAI whose runtime holds API keys sits in the same blast radius.

   Draw a reference architecture for a modern data team, throw darts at it, and every throw hits a CVSS 9.0 or higher this week.

   Action items

   • Patch Argo CD to ≥3.2.12/≥3.3.10 and rotate every Kubernetes Secret in namespaces it can read — including model-registry tokens, HF PATs, and cloud credentials
   • Audit Iceberg/Polaris catalog configurations: enforce explicit storage credential scoping and add write-path allowlisting for table metadata locations
   • Inventory all agent frameworks in use and pin versions; patch PraisonAI immediately if present; add CVE-feed subscriptions for LangChain, CrewAI, AutoGen
   • Patch NGINX across all inference gateways and audit rewrite-module usage in model routing configs

   Sources:LiteLLM landed in the KEV catalog this week · An Ollama endpoint exposed to the public internet · PraisonAI, an open-source multi-agent framework · SANS AtRisk newsletter