28 — Drift delta (2026-07-05)

The Volume IV drift chapter: on 2026-07-05 every drift-prone, load-bearing claim in the dossier was re-checked against live primary sources — platform/code/support docs, the Claude Code changelog, Anthropic announcements, provider pricing pages, GitHub issues, and arXiv. Volumes I–III stay frozen as written. Where a recorded claim is now false, the correction lives here; for the worst (load-bearing) cases an inline stale-claim marker was also added in the affected file (list in Corrections applied inline). Each area table lists all changes and only the materially interesting confirmations. Status vocabulary distinguishes honesty classes: CONFIRMED (claim verbatim-true today), DOCS CHANGED (claim was correct when recorded; the world moved), WAS WRONG (the recorded claim was incorrect on its own date), UNVERIFIABLE (source no longer states it either way). All sources in this chapter were accessed 2026-07-05; full URLs in the verification ledger.

TL;DR

• Claude Sonnet 5 (claude-sonnet-5) shipped 2026-06-30 and is the single biggest drift. Default model in Claude Code (2.1.197), 1M native context, 128k output, intro pricing $2/$10 per MTok through 2026-08-31 then $3/$15 — and it uses the new (Opus 4.7+) tokenizer. The dossier's compounded "price ratio × tokenizer ratio" downtier arithmetic (05, 10) now applies only to legacy Sonnet 4.6 / Haiku 4.5, not to the current mid tier. Also supersedes 24 — Fresh literature & market delta (clean-room re-sweep)'s "no post-06-12 change to API cost primitives" verdict.
• The Fable 5 subscription story was superseded twice. Access was suspended 2026-06-12→06-30 under a government export-control directive, then redeployed 2026-07-01; Pro/Max/Team/select Enterprise include Fable 5 up to 50% of weekly usage limits through 2026-07-07, after which it moves to usage credits at API rates. The 06-23-removal narrative in 19 — Subscription & quota economics / 24 — Fresh literature & market delta/27 is wrong during the July 1–7 window and reverts to credits-only after 07-07.
• The 2026-06-15 Agent-SDK/headless billing split was PAUSED by Anthropic — claude -p, SDK, and third-party usage still draw the interactive subscription cap. 19 — Subscription & quota economics Q7's "route CI off the cap" lever is currently inoperative.
• Beta/status drift, all in the loosening direction: the memory tool (memory_20250818) went GA (no beta header, 08); mid-conversation system messages are GA on Opus 4.8 (no beta header, 07); the context-windows-page thinking-retention contradiction recorded in 12 was fixed upstream; and Claude Code 2.1.198 made the built-in Explore agent inherit the main session's model instead of running on haiku — Explore's implicit cheap-model-routing saving (06) is gone.
• Docs bookkeeping moved under several recorded numbers without changing substance: count_tokens rate tiers renamed Start/Build/Scale at 2,000/4,000/8,000 RPM (the 100-RPM tier-1 row is gone; 01, 05); the pricing page's tokenizer note now says "approximately 30% more tokens" (the "up to 35%" phrasing is gone; 01); the limits doubling was announced 2026-05-06, not 2026-06-05 as recorded (19); and the token-counting docs now document claude-fable-5 as a directly supported count_tokens model id, staling the opus-4-8-proxy workaround (tools/18/20).
• The headline verdict stands: ≈2.5× defensible / ≈5–6.2× with validated routing / no honest 10× at equal quality, with tasks-per-cap as the subscriber metric. No delta touches either binding constraint (frontier thinking billed as output; the cache-read floor of genuinely-used context); Sonnet 5's intro pricing changes absolute dollars and which routing target is correct — the ≈5–6.2× routing path must now be pinned to old-tokenizer targets (Sonnet 4.6/Haiku 4.5) or re-derived at Sonnet 5's ratio — and the paused SDK split plus the Fable churn make the cap even more clearly the subscriber's binding constraint.

---

A. Pricing & models

All headline dollar numbers in the dossier remain correct: Fable 5 $10/$50 (cache $1/$12.50/$20, batch $5/$25), Opus 4.8 $5/$25, Sonnet 4.6 $3/$15, Haiku 4.5 $1/$5, batch −50% stacking with cache, 1M-context standard-priced, US data residency 1.1×, web search $10/1,000, code execution 1,550 free hrs then $0.05/hr — all confirmed on the live pricing page. The changes:

New facts the dossier predates: batch extended-output beta output-300k-2026-03-24 gives Opus 4.8/4.7/4.6, Sonnet 5, and Sonnet 4.6 up to 300k output tokens in the Batch API (synchronous stays 128k — 09/12); Fable 5's refusal/fallback billing surface is documented (stop_reason refusal at HTTP 200, no billing for pre-output refusals, server-side fallbacks beta, and "fallback credit" refunding the prompt-cache cost of a cross-model retry — changes the fallback-cost arithmetic in 07/10).

B. Caching

Cleanest area of the sweep: multipliers (0.1×/1.25×/2×), free TTL refresh, every min-cacheable-prefix floor, 4 breakpoints / 20-position lookback, max_tokens: 0 pre-warm with its full rejection list, workspace-scoped isolation (2026-02-05), the concurrency rule, all 8 Claude Code invalidators (with the v2.1.86 and v2.1.163 gates), the cache-safe list, the TTL policy (1h main loop on subscription / 5m subagents / 5m on credits and API, both env vars), /rewind past-TTL warmth, and excludeDynamicSections — all verified verbatim on the live pages. No load-bearing drift. The changes:

New: a dedicated Cache diagnostics docs page now exists for locating exactly where two requests diverged on an expected cache hit (07 technique 12 predates it); Claude Code gained a model-family-specific DISABLE_PROMPT_CACHING_FABLE env var; and Anthropic published a design-rationale blog ("Lessons from building Claude Code: Prompt caching is everything") citable for techniques 1/5/6 in 07 — Caching exploitation.

C. Claude Code changelog

The dossier's reference point was ~v2.1.175; latest is 2.1.201 (2026-07-03) — ~26 releases of drift, two of them load-bearing.

Other post-2.1.175 token-relevant releases: 2.1.198 subagents run in background by default and subagents+compaction inherit extended-thinking config (raises their output cost); 2.1.199/2.1.200 subagents cut off by rate limits/server errors return partial work (reduces wasted-token cost of failed runs); 2.1.181 fixed cache reads silently broken on custom ANTHROPIC_BASE_URL and Foundry (gateway users paid 1× miss rates before it); 2.1.191 self-compaction reminder for MEMORY.md; 2.1.178 TeamCreate/TeamDelete removed, spawn-time classifier gates team blowup (25); 2.1.196/2.1.187/2.1.175 org default models and model restrictions enforced across all pickers (27 governance lever).

D. Effort & thinking

All headline claims hold: effort ladder low→max, high the universal API default (effort: "high" ≡ omitting), still zero published per-level token percentages (the dossier's biggest-gap statement stands), Fable 5 thinking cannot be disabled, MAX_THINKING_TOKENS a no-op on Fable 5, budget_tokens deprecated (4.6-era) or 400 (4.7+), usage.output_tokens_details.thinking_tokens live, prior-turn thinking kept+billed on Opus 4.5+/Sonnet 4.6+, ultrathink prompt-level-only, ultracode = xhigh + orchestration permission, per-model effort calibration, billing invariant under thinking display modes. The changes:

New facts: Sonnet 5 supports the full effort ladder with a sanctioned cross-model anchor ("Sonnet 5 medium ≈ Sonnet 4.6 at high effort") and, unlike Fable 5, allows thinking: {type: "disabled"}; Mythos Preview exists (max but not xhigh, manual budget_tokens still accepted); Task budgets (beta) add an advisory token budget for the full agentic loop on Opus 4.7+/Fable 5/Mythos 5 — these models no longer receive the injected <budget:token_budget>/<system_warning> tags that Sonnet 5/4.6/4.5 and Haiku 4.5 get (candidate new record for 12, composability note for 09); CLAUDE_CODE_EFFORT_LEVEL accepts auto, orgs can cap effort levels per model, and unsupported levels fall back to the highest supported level at or below; on Fable 5, prompting for internal reasoning in response text can hard-refuse with stop_details.category: "reasoning_extraction" (affects any elicit-reasoning-in-text advice in 04); Fable 5/Mythos 5 are not available under zero-data-retention arrangements.

E. Context-management betas

Context editing (context-management-2025-06-27, clear_tool_uses_20250919 at 100k/keep-3, clear_thinking_20251015 first-in-array) and compaction (compact-2026-01-12, compact_20260112, 150k trigger / 50k minimum, same-model-only, usage.iterations billing shape, custom instructions replace the default prompt) all remain beta with unchanged defaults and types; still zero Claude Code exposure on any of the three pages. The changes:

New: the clear_thinking_20251015 default is now documented as model-specific (Opus 4.5+/Sonnet 4.6+ keep all prior thinking; earlier models keep last turn) — a second live page corroborating 12's retention record; the default compaction summarization prompt now "varies by model"; and the memory-tool docs ship the multi-session initializer→progress-log pattern that 08 technique 6 already describes, now doc-backed.

F. Tool-use features

All headline numbers in records 6/7/13 hold: both tool-search variant strings and limits (200-char regex, 10,000 deferred max, 30–50-tool degradation, ~55k example, "over 85 percent"), the cache-preservation and strict-mode-composition quotes verbatim, PTC's billing-exclusion quote + platform matrix + all four benchmark figures (+11%/24%, ~38%, 20–40%, τ²-bench −8%), container pricing, structured-outputs GA + 24h grammar cache + 20-strict-tools limit. The changes:

New: code_execution_20260521 (same runtime as _20260120, discloses the 90-second per-cell limit; interchangeable in allowed_callers) — record 7's "requires code_execution_20260120" should read "or later"; Sonnet 5 and Mythos 5 join the tool-search/PTC/structured-outputs matrices; tool search is ZDR-eligible while PTC is not, and on Bedrock server-side tool search works only via InvokeModel; a defer_loading: true tool cannot also carry cache_control (400) and at least one tool must stay non-deferred; PTC gained documented constraints (disable_parallel_tool_use: true unsupported, recursive $ref 400s, ~4-min pending-call timeout, ~5-min idle container reclaim); structured-outputs complexity limits itemized (20 strict tools, 24 optional params, 16 union-typed params per request).

G. Quota & subscription

Prices ($20/$17 Pro, $100/$200 Max), the 5-hour + weekly window structure, the Sonnet-only weekly limit, no Opus-specific cap, cross-surface pooling, 200K subscription context, usage credits at API rates with the $2,000/day redemption cap, the cache-read ITPM exclusion + 10% billing, the undocumented anthropic-ratelimit-unified-* family, and the silence on #24147 (no staff comment ever) and #45756 (none since 2026-04-12) all confirmed. The changes:

New: rate-limit tables now carry Fable 5 and Sonnet 5 as separate model classes (Fable 5: 1,000 RPM / 500K ITPM / 100K OTPM at Start tier; Opus is one combined 4.5–4.8 bucket); Haiku 3.5 — the sole cache-read-counts-toward-ITPM exception — is retired except Bedrock/Google Cloud, making the exclusion near-universal; fast mode has its own limits and a new anthropic-fast-* header family (second undocumented-to-dossier family beside unified-*); API monthly spend caps renamed and revalued Start $500 / Build $1,000 / Scale $200,000.

H. Cross-provider

All headline numbers hold: GPT-5.5 $5/$30 with $0.50 cached and 24h default retention, gpt-5.5-pro with no cached tier, reasoning_effort none→xhigh, no GPT-5.5 successor; Gemini implicit caching default on 2.5+ (2,048–4,096 minimums), explicit-cache ~0.1× reads + storage rent; Copilot 1 AI credit = $0.01, token billing at published per-model rates, Auto −10%, Fable 5/Opus 4.8 first-class rows. The changes:

New: the Gemini caching doc now targets the Interactions API, which supports implicit caching only — the explicit-cache + storage-rent lever narrows to the legacy generateContent surface (12/23); Gemini lineup advanced (3.5-flash GA $1.50/$9.00; 3.1-pro-preview $2/$12 ≤200k; 3.1-flash-lite $0.25/$1.50); Copilot GPT-5.5 gained a long-context tier ($10/$1/$45 vs default $5/$0.50/$30); Copilot included credits are Pro 1,500 / Pro+ 7,000 / Max 20,000 pooled at billing-entity level, with a legacy request-billed cohort ($0.04/request) persisting; OpenAI container sessions bill per-minute with a 5-minute minimum since 2026-06-02.

I. Fresh literature

Sweep window 2026-06-15→07-05: no new paper moves a dossier headline number. "Hold Onto That Thought" (2512.12008) and CAG (2412.15605) confirmed at their recorded claims; the KV-compression family verdict (self-host-only, $0 on hosted Claude) is reinforced by three new in-window members that all land inside the inference engine. One recorded claim drifted:

New in-window and oversight items worth folding in: InfoKV (2606.26875, entropy-aware KV compression for reasoning), CompressKV (2606.24467, >97% performance at 3% KV budget), MosaicKV (2607.00760, 2-D compression, 16× attention speedup) — all self-host, FL4 unchanged; FastContext (2606.14066, Microsoft, up to 60% token reduction + up to +5.5% SWE-bench resolution — currently withdrawn by author over product-IP concerns, cite only with that flag); "Brevity is the Soul of Inference Efficiency" (2606.25432, 35× Cost-of-Pass from curated concision, VLM — supports 09/20's direction); and two pre-window oversights uncited in 24: "Don't Break the Cache" (2601.06007 — 500+ agent sessions across three providers, caching cuts API cost 41–80% and TTFT 13–31%, strategic placement beats naive full-context caching; first peer-style empirical validation of the dossier's cache-first math, belongs in 07/24) and 2605.11051 (In-Context Autoencoder soft compression fails on multi-step agentic coding — direct reinforcement of FL3), plus TACO (2604.19572 — training-free rule-based filtering of terminal observations that gains 1–4% accuracy while cutting tokens; a counter-datapoint that nuances FL3's scope to model-based compressors, deterministic observation filtering can be negative-cost).

J. Tokenizer & counting

Confirmed: count_tokens free with an independent rate pool, estimate semantics, system-added tokens unbilled, prior-turn thinking ignored by the counter, never warms the cache, ZDR-eligible; the token-counting page's "roughly 30% more tokens" quote and the count-twice-across-the-4.7-boundary guidance verbatim; all recorded prices, so the 3.3× list / 5.3× effective ratios remain arithmetically valid for the old-family route. The changes:

The local family-equality measurements (Fable 5 ≡ Opus 4.8; Sonnet 4.6 ≡ Haiku 4.5) are consistent with the updated docs and need no change.

Corrections applied inline

The load-bearing findings above also received inline stale-claim markers in the affected files (applied by the sibling correction pass; this list is the authoritative index):

1. 10 — model routing & delegation — add Sonnet 5 to the baseline and effective-multiplier tables ($2/$10 intro → $3/$15; 1M context; 128k output; new tokenizer; effort default high); mark Sonnet 4.6 legacy with 128k max output; scope the tokenizer-ratio bonus to Sonnet 4.6/Haiku 4.5 only.
2. 24 — fresh literature & market delta — Fable 5 suspension 06-12→06-30, redeployment 07-01, 50%-of-weekly-limits plan inclusion through 07-07, credits after; Sonnet 5 exception to the "API primitives did not move" verdict and to the "tiering flips to Opus 4.8" delta.
3. 06 — context architecture — since 2.1.198 Explore inherits the main session's model (capped at opus) instead of haiku; subagents and compaction inherit extended-thinking config; subagents run in background by default.
4. 08 — retrieval, memory & state — memory tool is GA on the Messages API, no beta header (type still memory_20250818, all Claude 4+ models).
5. 19 — subscription & quota economics — the 2026-06-15 Agent-SDK split is paused; headless/claude -p/third-party usage still draws the interactive cap; Q7's lever inoperative until unpaused.
6. 05 — tokenizer arbitrage — new-family membership now includes Sonnet 5 (and Mythos Preview); all cross-family arbitrage numbers pinned explicitly to Sonnet 4.6/Haiku 4.5; Fable→Sonnet 5 carries no tokenizer discount.
7. tools hub, 18 — extension overview, 20 — multimodal token economics — the "count_tokens rejects claude-fable-5, use Opus 4.8 proxy" note is stale per docs; marked for one live-call drift check before final rewording.

Verification ledger

Every source consulted in this pass, all accessed 2026-07-05. Fetch failures: anthropic.com/news/redeploying-fable-5-mythos-5 (404 as linked from the introducing-fable-5 Tip; successor fetched), support.claude.com/en/articles/11049408 (404; successor 11049741 fetched), docs.github.com/en/copilot/concepts/billing/models-and-pricing (404; moved under reference/copilot-billing), and two WebFetch truncations on the very long compaction page (resolved via the raw .md variant).