Anthropic Quietly Cut Cache-Read Prices to a Quarter, Turned on Adaptive Thinking by Default, and Shipped Claude Fable 5.1 Across Five Clouds

<h1>Anthropic Quietly Cut Cache-Read Prices to a Quarter, Turned on Adaptive Thinking by Default, and Shipped Claude Fable 5.1 Across Five Clouds</h1>

<p>Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026 with a 1M-token context, 128k max output, always-on adaptive thinking, the same $10 / $50 per million input/output token pricing as Claude Fable 5, and cache reads priced at $0.25 per million tokens — down from the $1/MTok Anthropic charged on every prior model. Claude Fable 5.1 is available immediately on the Claude API plus Bedrock, Claude Platform on AWS, Google Cloud (Vertex AI), and Microsoft Foundry. The Mythos tier is gated to Project Glasswing participants.</p>

<p>Three breaking changes ship with the launch and matter immediately if you write the messages array yourself. Five additive features target long-running autonomous agents — per-message effort changes, turn-scoped system messages, between-tool-call progress updates, the 75% cache-read cost cut, and content provenance (text watermark + C2PA on images and video produced through the code-execution + Files API pipeline). If you run Claude Code v2.1.257+ with Fable 5.1 as the default, or if you operate Bedrock / Vertex AI / Foundry workloads against the Fable 5 pricing family, today is the migration day.</p>

<h2>What happened</h2>

<p>Anthropic's Claude Platform release notes for September 1, 2026 state the launch verbatim:</p>

<blockquote>We've launched Claude Fable 5.1 (claude-fable-5-1), the successor to Claude Fable 5 for long-running agentic coding, knowledge work, and research, alongside Claude Mythos 5.1 (claude-mythos-5-1) for Project Glasswing participants. Both models support a 1M token context window by default, 128k max output tokens, and always-on adaptive thinking, at $10 / $50 USD per MTok, the same as Claude Fable 5, with cache reads cut to $0.25 per MTok.</blockquote>

<p>A second bullet in the same release confirms the pricing treatment and clarifies the cache multiplier change: cache reads on Fable 5.1 and Mythos 5.1 cost $0.25 USD per million tokens — 0.025x the base input price, compared with 0.1x on other models. Cache writes are unchanged. The pricing footnote on the Anthropic pricing page reads: "Cache hits and refreshes on Claude Fable 5.1 and Claude Mythos 5.1 are priced at 0.025x the base input price. All other models use the standard 0.1x multiplier."</p>

<p>Fable 5.1 is the documented default in Claude Code v2.1.257 and later. Mythos 5.1 is restricted to Project Glasswing participants — Anthropic's gated, application-only access tier for the Mythos preview line. Don't plan on Mythos for production today unless you've already applied and been admitted.</p>

<h2>What actually changed</h2>

<p>The same release-notes bullet above lists eight distinct surfaces that need review if you operate agent stacks. Six are additive; three are breaking.</p>

<p>Breaking — changes that will reject requests today</p>

<li>Forced tool use is removed. tool_choice: {"type": "any"} and tool_choice: {"type": "tool", "name": "..."} return a 400 invalid_request_error. The default auto and none are unchanged. Anthropic's stated reason: thinking is always on, and a forced tool call would skip thinking and dump working-out into the tool arguments, lowering argument quality. Workaround: keep auto, set strict: true on the tool definition (strict tool use), or use structured outputs. To make the model call a specific tool when appropriate, say so in the prompt ("Use the get_weather tool to answer"). The [What's new in Claude Fable 5.1](https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1) doc confirms Fable 5.1 follows explicit tool instructions reliably.</li>

<li>Thinking blocks are tied to the producing model. Every thinking block records which model produced it, and preservation flows one direction only: a conversation that moves onto Fable 5.1 (from Opus 5, Fable 5, or earlier) keeps its reasoning. A conversation that moves off Fable 5.1 to any earlier model loses thinking for the turns that run there. Dropped blocks don't count toward input_tokens and aren't billed. Opt into thinking-binding-controls-2026-08-01 to receive dropped-block reports in a top-level input_transformations array; without it, the drop is silent.</li>

<li>Editing earlier turns invalidates thinking blocks. Modifying anything before a Fable 5.1 thinking block (the system prompt, tools, or an earlier message) returns a 400 on the next request or, with thinking-binding-controls-2026-08-01 and prefix_mismatch_behavior: "drop_block", silently drops the now-invalid blocks. The check is enforced for new accounts created on or after August 31, 2026; older accounts see the check only when you opt into the header. The migration guide's three-step check is prefix_mismatch_behavior: "drop_block" for a session, then log input_transformations, then inspect what got dropped.</li>

<p>Additive — capabilities you can flip on without code rewrites</p>

<li>Cache reads at $0.25/MTok. 0.025x base input. On every prior model the multiplier is 0.1x — so Fable 5.1 cache reads are 4x cheaper than Opus 5 cache reads ($0.50/MTok) and 4x cheaper than Fable 5 / Mythos 5 cache reads ($1.00/MTok). Cache writes are unchanged at 1.25x for 5-minute caches and 2x for 1-hour caches.</li>

<li>Per-message effort in beta (mid-conversation-output-config-2026-07-01 header). Add a role: "system" message with empty content and an output_config.effort field inside messages to change effort for the next turn while preserving the prompt cache. Raise it for the hard step, drop it for the routine ones. Supported on Fable 5.1, Mythos 5.1, and Opus 5.</li>

<li>Turn-scoped system messages in beta (mid-conversation-system-clear-at-2026-08-21 header). clear_at: "next_user_message" renders the message for the current turn, then it stays in history at zero token cost. Per-turn reminders no longer accumulate, don't invalidate the prompt cache, and don't invalidate later thinking blocks.</li>

<li>Readable progress updates between tool calls in beta (thinking-display-updates-2026-08-18 header). thinking.display: "updates" returns an empty thinking field but the short progress updates Fable 5.1, Mythos 5.1, and Fable 5 write between tool calls come back as text, with at most one thinking block before a tool call.</li>

<li>Content provenance. Text generated by Fable 5.1 / Mythos 5.1 carries Anthropic's text watermark. Supported image and video files produced through the code-execution tool carry C2PA Content Credentials when retrieved through the Files API. Marking requires no request or response-handling changes.</li>

<p>Token economics shift</p>

<p>The same tokenizer introduced with Opus 4.7 is used — Claude Fable 5.1 produces roughly 30% more tokens for the same text compared with Sonnet 4.6 and earlier. If your cost model assumes Sonnet-token counts but routes to a 4.7+ model, your effective per-request cost is roughly 1.3x the headline price. The token-counting endpoint applies the model-specific tokenizer.</p>

<p>Data retention requirement</p>

<p>Like Fable 5, both models require 30-day data retention and aren't available under zero data retention unless expressly authorized by Anthropic. If you've been hitting ZDR API limits for compliance reasons, those don't relax here.</p>

<h2>Why developers and founders should care</h2>

<p>If you run Claude Code for long-horizon coding work, Fable 5.1's adaptive-thinking-by-default + 1M context changes the right-bucket decision. Per the What's new page, Anthropic's model-selection guidance for autonomous coding work is now: "Use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short." Opus 5 remains the default for most workloads. Fable 5.1 is the swing model — not the new floor.</p>

<p>If you cache aggressively, 75% is not a small cut. A long-running autonomous agent that keeps a 600k-token context warm (the kind of context Claude Code v2.1.256+ produced with the SubagentModel forcing flag) at $0.25/MTok cache reads vs $1.00/MTok is roughly a 4x effective per-token cost reduction. For a workload consuming 200M cached input tokens per day at the new rate, that's $50/day vs $200/day at the old rate — a $150/day / $4,500/month delta per long-running agent. If you're running multiple agents in parallel (the pattern Cursor Cloud Agents subscriptions shipped last week and Claude Managed Agents deployments push you toward), the cumulative cache-read spend is where the new tier earns its place.</p>

<p>If you operate across clouds, the same model and pricing surface lands on all five host paths today: Claude API, Bedrock, Claude Platform on AWS (CCU billing through AWS Marketplace), Google Cloud / Vertex AI, and Microsoft Foundry (CCU billing through Azure Marketplace). If you've been held back from Bedrock or Vertex by a need for the latest tier, that's gone. Note the regional endpoint premium: for 4.5+ models on Bedrock and Google Cloud, regional and multi-region endpoints add a 10% premium over global endpoints. Affects your cost comparison only if you opt in to regional pinning.</p>

<p>If you depend on forced tool calls, your code today rejects a request that worked yesterday. The fix is mechanical — tool_choice: {"type": "tool", "name": "x"} becomes a prompt instruction to use tool x, plus strict: true on the tool definition. The validation hits the token-counting endpoint too, so don't be surprised when your pre-flight sizing call rejects before the actual request does.</p>

<p>If you build multi-turn agent scaffolds that mutate conversation history, the prefix-binding check is the change most likely to surface as silent 400s. The patterns that invalidate thinking blocks (per the What's new doc): editing/reordering/removing an earlier turn while keeping later ones; injecting per-request text into earlier turns (status lines, reminders) you remove on the next request; rebuilding top-level system or tools between requests in the same conversation; images or document URLs that serve different bytes on later requests. The patterns that keep blocks valid: trimming a leading run of thinking blocks (oldest first), letting server-side compaction or context editing trim, moving cache_control markers, changing effort. Claude Code, claude.ai, Claude Managed Agents, and the Agent SDK keep the prefix intact for you — direct API consumers with custom messages builders need to audit.</p>

<h2>Evidence and test results</h2>

<p>Primary sources verified at 2026-09-02 12:16 UTC:</p>

<li>Anthropic Platform release notes, September 1, 2026 entry. Verified verbatim against the live page. Eight bullets in the Fable 5.1 / Mythos 5.1 entry cover launch, cache pricing, tool_choice validation, preserved-thinking rules, per-message effort beta, turn-scoped system messages beta, progress updates beta, content provenance, and the 30-day data retention requirement.</li>

<li>Anthropic pricing page. Verified verbatim against the live page. Pricing table now lists five distinct cache columns per row: Base input, 5m cache writes, 1h cache writes, Cache hits and refreshes, Output tokens. Fable 5.1 and Mythos 5.1 rows show $0.25 cache reads; Fable 5 and Mythos 5 rows show $1 cache reads. The footnote confirms the 0.025x vs 0.1x multiplier distinction. Cloud platform pricing section unchanged on Bedrock and Google Cloud; Claude Platform on AWS and Microsoft Foundry sections bill via Claude Consumption Units (CCUs) at $0.01 per CCU after rate conversion.</li>

<li>What's new in Claude Fable 5.1. Confirmed verbatim. Three breaking changes and five additive features catalogued. Models table shows the API IDs and availability. The pricing section confirms the only model-rate change is the lower cache read price; input and output rates match Fable 5.</li>

<li>Claude Code release v2.1.257 (referenced in yesterday's pillar, claude-code-2-1-257-fable-51-containment-escape-rule-sep-2026) — Fable 5.1 as the new default with $10/$50 and $0.25 cache reads.</li>

<p>Documentation comparison only. No firsthand API call was run for this article. All pricing, validation, and beta-header claims are surface-level documentation that can be reproduced end-to-end by anyone with a Claude API key and claude-fable-5-1 access. Test workloads against the per-message effort beta and prefix-binding check should be expected to surface silent-drop behavior depending on account age (post-2026-08-31 accounts enforce; pre-2026-08-31 accounts report only when the header is set).</p>

<p>Independently corroborating evidence. Yesterday's pillar article confirmed Fable 5.1's role as the new default in Claude Code v2.1.257, the $10/$50/$0.25 pricing line, and the seven security/sandbox fixes in that release. Today's piece complements, doesn't replace, that article: yesterday's was the Claude Code product-and-security sweep; today's is the platform-level capability surface and cost shift.</p>

<h2>Cost, risk, and limitations</h2>

<p>Cost is the headline win. On a cached-heavy workload the cache-read cost cut is the single biggest line-item change. Four example workloads, all assuming 200M cached input tokens / day at $0.25/MTok:</p>

<p>These are cache-read savings only. Output tokens and non-cached input tokens still cost the same. If your workload is output-heavy or has a cold cache, the new tier doesn't move the needle. If your workload is the canonical agentic-coding loop — repeated cache hits across turns — the new tier is a 4x reduction on the cached portion.</p>

<p>Risk profile (numbered against Anthropic's release notes):</p>

<li>Forced tool use is gone. Code paths relying on {"type": "tool"} will 400. Audit your prompt-engineering code for that pattern; if you find it, plan a migration to strict: true plus prompt-driven tool instruction.</li>

<li>Thinking-block migration cost. Code that mutates prior turns (status lines, doc-URL updates, system-prompt rebuilds) will lose reasoning or return 400s. The migration path is explicit: switch to mid-conversation system messages with clear_at for turn-scoped reminders, use server-side context editing / compaction for context trimming, audit your messages builder.</li>

<li>30-day data retention is required. If you're a regulated workload (HIPAA, FedRAMP, GDPR with strict data-residency controls), the ZDR non-availability for Fable 5.1 may push you to Sonnet 5 at the existing $2/$10 pricing.</li>

<li>Token-count inflation. The 4.7+ tokenizer produces ~30% more tokens for the same text. Your cached-context size in tokens is roughly 30% larger than an equivalent Sonnet 4.6 context. Multiply your per-token budgets accordingly.</li>

<li>Beta-header surface. Per-message effort, turn-scoped system messages, and progress updates between tool calls are all beta features gated behind specific beta headers. Production-grade adoption should expect the headers to evolve before they go GA.</li>

<li>Mythos 5.1 is gated. Don't design production systems around Mythos unless you're already admitted to Project Glasswing.</li>

<p>What we did NOT test:</p>

<ul><li>No firsthand API call against claude-fable-5-1 was run for this article. All claims are documentation-derived.</li></ul> <ul><li>The 30-day data retention enforcement boundary at the API level was not tested against actual account configurations.</li></ul> <ul><li>Beta header behavior (mid-conversation-output-config-2026-07-01, mid-conversation-system-clear-at-2026-08-21, thinking-display-updates-2026-08-18) was not exercised.</li></ul> <ul><li>C2PA marking behavior on code-execution-generated files was not validated.</li></ul> <ul><li>Account-age behavior on the prefix-binding check (post-2026-08-31 enforcement vs older-account opt-in) was not empirically confirmed.</li></ul>

<h2>Mr. Technology verdict</h2>

<p>This is the cleanest documentation-comparison shape we've seen all week: eight documented surfaces, primary sources for the launch announcement, the pricing table, and the "what's new" reference page, all cross-referenced. The headline numbers — 4x cache-read cost reduction, 1M context default, adaptive-thinking-by-default — are credible on inspection.</p>

<p>The breaking-change surface is contained but real. Three of the eight documented changes reject requests today. If you write the messages array yourself, this is a migration day; if you run Claude Code, Claude Agent SDK, Claude Managed Agents, claude.ai, or any of Anthropic's first-party SDKs, Anthropic's documentation indicates the prefix-binding protection is handled for you.</p>

<p>This is publishable as a pillar article because:</p>

<ul><li>The cost reduction is large enough to change deployment economics for any operator running more than one cache-heavy agent (a near-universal pattern among developers/founders running agents).</li></ul> <ul><li>Three breaking changes affect direct-API users; the migration matrix is short and explicit.</li></ul> <ul><li>One new model tier (Mythos 5.1) is gated, but the Fable 5.1 tier is generally available across five clouds.</li></ul> <ul><li>The launch event is timestamped, primary-sourced, and verifiable end-to-end with a Claude API key.</li></ul> <ul><li>The 75% cache-read cut is the kind of concrete number that can be acted on in a budget review, not just noted.</li></ul>

<p>Article-level caveats that warrant re-verification on a subsequent run:</p>

<ul><li>The token inflation figure (30%) is from Anthropic's own docs; no independent benchmark against representative agent traces was run.</li></ul> <ul><li>Beta-header surfaces may change shape before GA — the per-message effort header in particular has a 2026-07-01 beta suffix that suggests an earlier preview.</li></ul> <ul><li>The Cloud Platform on AWS / Microsoft Foundry billing surface via CCUs at $0.01/CCU is unchanged; rate-conversion accuracy was not independently audited.</li></ul>

<h2>Recommended action</h2>

<p>Today (September 2, 2026), do this:</p>

<ul><li>For Claude Code users: confirm your install is v2.1.257+ so Fable 5.1 is the default. The claude --help line for model selection will show Fable 5.1. If you've manually pinned the model, your pinned model still wins.</li></ul> <ul><li>For Bedrock / Vertex AI / Foundry operators: confirm Fable 5.1 is the new default in your request templates by checking the live model list. If you've been on Fable 5 with regional endpoints and a 10% premium, decide whether the cache-cut justifies flipping to Fable 5.1 today or after a measured ramp.</li></ul> <ul><li>For direct API users with custom messages builders: audit for tool_choice: {"type": "any"} and tool_choice: {"type": "tool", ...} patterns. Both now 400. Plan a migration to auto + strict: true on the tool definition, plus a prompt-level "use tool x to answer" instruction.</li></ul> <ul><li>For multi-turn agent scaffolds that mutate prior turns: switch to mid-conversation system messages with clear_at: "next_user_message" for turn-scoped reminders. Use server-side context editing / compaction for context trimming. Avoid rebuilding the top-level system or tools array between requests in the same conversation.</li></ul> <ul><li>For budget planning: rebuild your monthly cost model with the 0.025x cache-read multiplier. Cache-heavy workloads (re-reading large system prompts, agent loops over tool results, multi-turn coding sessions with shared context) will see a 4x reduction on the cached portion of input spend. Don't assume the saving is symmetric; output tokens and non-cached inputs are unchanged.</li></ul>

<p>This week:</p>

<ul><li>Test the per-message effort beta on a representative workload. The cost/quality tradeoff of high-effort for hard steps and low-effort for routine ones is worth measuring against your existing baseline before adopting.</li></ul> <ul><li>Test the prefix-binding check on your longest-running conversation pattern. Use prefix_mismatch_behavior: "drop_block" with thinking-binding-controls-2026-08-01 for one session, log input_transformations, and confirm no unexpected mutations are happening upstream of your reasoning.</li></ul> <ul><li>Pin a C2PA validation pass on code-execution-produced images and video if you have a downstream content-provenance workflow.</li></ul>

<p>Skip if not in scope:</p>

<ul><li>Don't migrate to Fable 5.1 if your workload is output-heavy and cache-cold. The 4x cut on cache reads doesn't move the dial.</li></ul> <ul><li>Don't plan production use of Mythos 5.1 unless you're a Project Glasswing participant.</li></ul> <ul><li>Don't disable zero-data-retention enforcement to access Fable 5.1. The 30-day data-retention requirement is a security/regulatory control, not a marketing footnote.</li></ul>

<h2>Sources</h2>

<ul><li>Anthropic Platform release notes — September 1, 2026 entry (verified 2026-09-02 12:16 UTC)</li></ul> <ul><li>Anthropic pricing page (verified 2026-09-02 12:16 UTC; table and footnote quoted verbatim)</li></ul> <ul><li>What's new in Claude Fable 5.1 (verified 2026-09-02 12:16 UTC)</li></ul> <ul><li>Project Glasswing (Mythos availability tier)</li></ul> <ul><li>Claude Code 2.1.257 release notes (referenced for default-model confirmation)</li></ul> <ul><li>Mr. Technology — Claude Code v2.1.257 article (Sep 1, 2026) (yesterday's complementary piece)</li></ul>

Originally published: 2026-09-02 12:08 UTC Last verified: 2026-09-03 16:13 UTC No corrections at this time.