AlphaSignal - Anthropic doubles Claude token limits, Hermes Agent v0.14, Cursor Composer 2.5 (May 19 2026)

Why this is in the vault

Anthropic doubling Claude token limits across every paid plan (and a 1500% Tier-1 input-tokens-per-minute API rate-limit bump) materially changes the cost-of-context calculation that drives CLAUDE.md hard rule #4 and the agentmemory paper-trade queued from yesterday's AlphaSignal note - the ceiling moved, but the floor (context rot) didn't.

Issue contents

Top item: Anthropic doubles Claude token limits across all plans (13.5k likes)

• Doubled token limits on every paid plan; "peak-hour limits removed" for Pro and Max.
• API Tier-1 input-tokens-per-minute reportedly up ~1500%.
• Lior frames the underlying enabler as Anthropic's new SpaceX/Colossus deal granting access to 220,000+ Nvidia GPUs.
• RDCO relevance: strong. Doubled per-session ceiling reduces the cost of routing-mistakes against hard rule #4 but doesn't repeal it - "context rot" is a model-performance curve, not a hard cap. See Kepler sponsor blurb below for the same point from a third party.

Top item: Cursor ships Composer 2.5 (17.2k likes)

• Same open-source Kimi K2.5 base as 2.0; trained on 25x more synthetic tasks with mid-task feedback.
• 79.8% SWE-Bench Multilingual (up from 73.7%); claims parity with Claude Opus 4.7 and GPT-5.5 on coding benchmarks, "under $1 per task" vs competitor $11.
• IDE / CLI / web only - no public API.
• Bigger signal: Cursor is training a from-scratch model with SpaceXAI on a 1M H100-equivalent cluster (separate project).
• RDCO relevance: medium. Not switching IDEs from Claude Code; tracking as evidence the harness+model-trained-for-harness flywheel is real. SpaceXAI cluster appearing in both this item and the Anthropic item is the curation-level story Lior pulled out.

Top item: Nous Research ships Hermes Agent v0.14.0 (4.6k likes)

• New OpenAI-compatible local proxy converts OAuth-authed subscriptions (Claude Pro, ChatGPT Pro, SuperGrok) into endpoints Codex/Aider/Cline/Continue can hit - "no more paying twice."
• xAI Grok via SuperGrok OAuth; grok-4.3 bumped to 1M token context.
• LSP diagnostics on every file write; /handoff transfers full session live to a different model.
• 22 messaging platforms total (LINE + SimpleX added).
• RDCO relevance: medium. The OAuth-proxy pattern is the interesting bit - it's the harness equivalent of arbitraging subscription vs API pricing. Not adopting (security-review default + Ray runs on Claude Code natively), but the pattern is worth noting as evidence the "harness as compute-economics router" thesis is materializing.

Signals (one-liners)

1. OpenAI Codex update with 10-50x faster Git ops + custom shortcuts. Skip-but-note - harness-primitive speed compounds.
2. Buildkite paid placement. Skip.
3. Open-source RAG library cuts storage 97% privately (11.4k stars). Skip-but-note - adjacent to yesterday's agentmemory.
4. Harvard/MIT red-team study: live agents "lie, destroy, and take over." Worth deep-fetch later - agent-safety-in-the-wild thesis surface, queuing to research backlog (not deep-fetching here per 2-link cap).
5. Qwen3 35B MoE quantized GGUF + speculative decoding. Duplicate of yesterday's signal. Skip.
6. Open-source tool cuts AI code exploration tool calls 92% via semantic graphs. Skip-but-note - second "92% fewer" curation-genre headline in two days; watch for inflation.

⚠️ Sponsorship

• FriendliAI - paid inference-optimization guide. Same placement as yesterday; framing is TTFT/TPOT/E2E latency measurement + quantization + speculative-decoding tuning. Standard disclosure, no RDCO action.
• Kepler - paid placement; "deterministic infrastructure for AI" framing, hiring NYC founding engineers. Same pitch as yesterday, same notable line worth quoting verbatim: "Past 200k tokens, sourcing discipline erodes. Bigger context windows don't fix sourced reasoning. They break it sooner." That line sits adjacent to the Anthropic token-doubling top item in the same issue - the implicit editorial juxtaposition is that the new ceiling doesn't repeal context rot. Sponsor commercial intent is hiring + infra-positioning, no immediate RDCO action.
• Buildkite - paid CI orchestration placement, name-drops big-lab customers. Skip.

All three are standard AlphaSignal paid placements, disclosed. Same sponsor mix as 2026-05-18 issue - AlphaSignal's sponsor stack appears stable week-over-week.

No alphasignal.ai self-cross-promo links beyond standard masthead nav (Signup / Work With Us / Follow on X / Archive); these are subscription/contact, not editorial self-promo.

Mapping against Ray Data Co

• CLAUDE.md hard rule #4 (route long artifacts through subagents): doubled per-plan token limits + 1500% API rate-limit bump softens the cost of an in-context mistake but doesn't eliminate it. The rule is about model performance degradation across long context, not a hard cap. Kepler's "past 200k tokens, sourcing discipline erodes" line is third-party corroboration. No change to rule #4.
• agentmemory paper-trade ([[2026-05-18-alphasignal-agentmemory-92-percent-fewer-tokens]]): doubled ceiling reduces the urgency of agentmemory adoption but doesn't change the direction. Across-session memory is still the right substrate question. Paper-trade still scheduled.
• Harness thesis arc: Cursor training a model specifically for its harness (Composer 2.5 on synthetic tasks with mid-task feedback) is one more data point that the model+harness coupling is real. Connects to the harness-engineering reference cluster.
• SpaceXAI / Colossus appearing in both Cursor and Anthropic top items: the compute-access-as-moat point Lior leads with is worth filing as a vault concept. Not promoting to a concept doc tonight; flagging for the curiosity backlog.
• Cursor Composer 2.5 model availability: IDE/CLI/web only, no public API. Means it cannot be wired into Ray's Claude Code harness as a model provider - hard pass on adoption regardless of benchmark numbers. Skip.
• Hermes OAuth proxy pattern: interesting as an arbitrage on subscription-vs-API pricing, but security-review default + Ray runs natively on Claude Code mean this is observation-only. Skip-but-note.

Verdict: medium mapping overall. Anthropic token doubling is the load-bearing item; everything else is observation. Skill-bundle paper-trade queued yesterday is unchanged. No new Notion tasks required - the Harvard/MIT agent red-team study is the one item worth eventually deep-fetching but not tonight.

Related

• [[2026-05-18-alphasignal-agentmemory-92-percent-fewer-tokens]] - yesterday's issue; sponsor stack is identical, Qwen3 35B signal duplicates
• [[2026-04-15-thariq-claude-code-session-management-1m-context]] - "more context isn't free" is the same point as Kepler's sponsor line
• [[2026-04-12-cobus-greyling-harness-era-language-shift]] - Cursor-trained-on-harness-feedback is one more harness-era data point
• [[2026-04-12-harrison-chase-harness-blog]] - "Your Harness, Your Memory" - the doubled ceiling doesn't change the memory-slot question
• [[2026-05-10-addy-osmani-agent-harness-engineering]] - most recent harness-engineering reference
• [[2026-04-13-jaya-gupta-ai-lock-in-state-moat]] - compute-access-as-moat (SpaceX/Colossus angle) is adjacent to state-as-moat