Claude Fable 5.1 Cuts Cache Reads 75% as Copilot Retires Older Models

Top Stories

Claude Fable 5.1 ships 75% cheaper cache reads for long-horizon coding agents

Yesterday’s brief called September 1 a control-plane day. The frontier model landed anyway. Anthropic released Claude Fable 5.1 and invite-only Claude Mythos 5.1 — same weights, different safeguards — with input/output still at $10 / $50 per million tokens and cache reads cut from $1 to $0.25. Anthropic’s own usage mix from August puts that at about 25% less for typical Fable traffic and up to about 45% less for tool-heavy agent runs. Model id is `claude-fable-5-1` on the Claude API; Bedrock is `anthropic.claude-fable-5-1`. Context is 1M by default, max output 128K, adaptive thinking always on, default effort `high`, knowledge cutoff June 2026. Cache minimum on Fable 5.1 is 512 tokens.

Use Fable 5.1 when Opus 5 at high effort still falls short on long agent loops. Do not treat it as a daily-driver swap for Sonnet 5 ($2 / $10) or Opus 5 ($5 / $25). Three breaking API changes vs Fable 5: forced tool use errors, earlier models cannot read Fable 5.1 thinking blocks, and editing earlier turns invalidates thinking. Fable 5.1 still routes dual-use cyber (pentest, exploit generation, binary vuln scanning) to Opus 4.8 and research-biology queries to Opus 5; Anthropic says Claude Code sessions should see about 60% fewer cyber-safeguard interventions than Fable 5 because the model may now identify source-code vulnerabilities. Mythos 5.1 stays behind Project Glasswing / CVP / LSVP. Enterprise Frontier Safeguards (customer-controlled storage) start rolling out this fall; eligible customers keep ZDR on Fable 5.1 until then. Official announcement, model spec, and prompt-caching price table: Anthropic, Fable 5.1 overview, prompt caching.

Claude Fable 5.1 is generally available in GitHub Copilot, off by default for Business and Enterprise

GitHub put Fable 5.1 on the Copilot model picker in VS Code, Visual Studio, Copilot CLI, coding agent, the Copilot app, github.com, GitHub Mobile, JetBrains, Xcode, and Eclipse. Access is Pro+, Max, Business, and Enterprise. Copilot Business/Enterprise admins must flip the Claude Fable 5.1 policy in Copilot settings; it ships off. Rollout is gradual. Billing is provider list pricing under usage-based Copilot credits (1 credit = $0.01). GitHub’s model table lists Fable 5.1 at $10 input / $0.25 cache read / $12.50 cache write / $50 output per million tokens — the same 75% cache-read cut as the API, versus $1.00 cache reads on Fable 5.

The policy ack is the real gate. Fable 5.1 (and Fable 5) require Anthropic data retention by default so safety classifiers can run; retained prompts/outputs are not used for training. Other Claude models in Copilot stay on zero data retention. Eligible enterprises can get ZDR through the end of the calendar year while EFS rolls out; GitHub Support cannot grant that — contact the account team, then still enable the policy. Enabling the policy is the ack. Changelog and pricing: GitHub Changelog, Copilot models and pricing, supported models.

GitHub Copilot deprecates Gemini 3.1 Pro, Claude Opus 4.5/4.6, Sonnet 4.5/4.6, and Raptor Mini

Effective 2026-09-01 (changelog dated 2026-08-31), GitHub removed six models from most Copilot surfaces — Chat, inline edits, ask and agent modes, and completions:

Claude Sonnet 4.6 remains for individual Copilot subscribers on annual plans. If you pinned any of the retired ids in VS Code `github.copilot.chat.model`, agent workflows, or org model policy, those selectors go empty. Copilot Enterprise admins need to enable the replacement models in Copilot settings before they appear in VS Code or on github.com. Nothing is required to delete the old ones. Deprecation changelog.

Claude Code 2.1.257–2.1.258 makes Fable 5.1 default and tightens auto-mode permissions

Claude Code 2.1.257 (2026-09-01) adds `claude-fable-5-1` as the default Fable model: 1M context, $10/$50, $0.25/MTok cache reads. The `fable` alias now resolves there on first-party traffic. Exception: Claude apps gateway sessions keep resolving `fable` and `best` to Fable 5 until those gateways accept 5.1; pick Fable 5.1 in `/model`. npm `@anthropic-ai/claude-code@2.1.258` is current and fixes a 2.1.255 macOS 12 (Monterey) launch regression.

Security changes matter more than the model alias if you run auto mode or CI. 2.1.257 adds a Containment Escape rule so cloud-metadata credential fetches, egress evasion, and cross-tenant reach are not auto-approved unless the environment marks them expected; auto mode now prompts once before the first read outside the working directories (`permissions.blockReadsOutsideWorkingDirectories` to refuse). `defaultMode: “bypassPermissions”` in `.claude/settings.json` or `.claude/settings.local.json` is ignored (same as `”auto”`); set it in user or managed settings, or pass `–permission-mode`. `–add-dir` / `additionalDirectories` now refuse UNC and `/net/` paths. Several permission-rule bypasses closed: `permissions.ask` skipped inside compound commands/subshells, Bash `Read()`/`Edit()` denies missing `< file` redirects and `tac`/`egrep`, trailing-dot sandbox hosts evading `deniedDomains`, and `--disallowedTools` dropped after the first settings reload under `allowManagedPermissionRulesOnly`. Credential leaks to Foundry/Vertex/Bedrock on double `Authorization` headers are fixed. Changelog: Claude Code, npm 2.1.258.

Cursor Origin API adds repository.deleted and stops omitting default-valued REST fields

Origin is still alpha, but Sep 1 is a client-break if you treat missing JSON keys as defaults. REST responses now include fields at their default value: a non-draft PR reports `”draft”: false` instead of omitting `draft`; empty lists are `[]`. Optional-when-unset fields such as `submitted_at` still disappear. That matches how Origin webhooks already serialized. New webhooks: `repository.deleted` (native/outbound delete, or stopping an inbound GitHub mirror — GitHub itself sends nothing for the Cursor-side repo) and `repository.metadata.updated` on default-branch changes. Both need `repository:metadata:read`. Requested-reviewer users now carry `email` next to `id`. OpenAPI `operationId`s are unique; dual-URL ops take a `_2` suffix (`OriginService_GetRepoTarball_2`, `OriginService_ListMatchingGitRefs_2`) — regenerate generated clients. `installation.updated` also fires on owner-namespace rename. Aug 31 added `x-origin-scopes` on every operation (`scopes` + `tokenTypes`: `app` / `installation` / `user`). Origin API changelog.

Tutorial

# --- bash ---
: "${ANTHROPIC_API_KEY:?export ANTHROPIC_API_KEY to your Anthropic Console key first}"
command -v jq >/dev/null || { echo "install jq"; exit 1; }
command -v python3 >/dev/null || { echo "install python3"; exit 1; }

# --- bash ---
python3 - <<'PY'
from pathlib import Path
block = (
    "You are a senior staff engineer reviewing a TypeScript monorepo. "
    "Prefer root-cause fixes over symptom patches. Do not invent APIs. "
    "Keep diffs minimal. Use the repo's existing lint and test scripts. "
    "Never print secrets. Cite file paths in repo-relative form. "
)
system = (block * 40).strip()

... click "Show full code" below to expand

# --- bash ---
: "${ANTHROPIC_API_KEY:?export ANTHROPIC_API_KEY to your Anthropic Console key first}"
command -v jq >/dev/null || { echo "install jq"; exit 1; }
command -v python3 >/dev/null || { echo "install python3"; exit 1; }

# --- bash ---
python3 - <<'PY'
from pathlib import Path
block = (
    "You are a senior staff engineer reviewing a TypeScript monorepo. "
    "Prefer root-cause fixes over symptom patches. Do not invent APIs. "
    "Keep diffs minimal. Use the repo's existing lint and test scripts. "
    "Never print secrets. Cite file paths in repo-relative form. "
)
system = (block * 40).strip()
Path("/tmp/fable51-system.txt").write_text(system)
print(f"system_words={len(system.split())} chars={len(system)}")
PY

SYSTEM="$(cat /tmp/fable51-system.txt)"
python3 - "$SYSTEM" <<'PY'
import json, sys
system = sys.argv[1]

def body(user, max_tokens):
    return {
        "model": "claude-fable-5-1",
        "max_tokens": max_tokens,
        "system": [
            {
                "type": "text",
                "text": system,
                "cache_control": {"type": "ephemeral", "ttl": "5m"},
            }
        ],
        "messages": [{"role": "user", "content": user}],
    }

# max_tokens 0 writes the cache and returns no output.
open("/tmp/fable51-warm.json", "w").write(json.dumps(body("warmup", 0)))
open("/tmp/fable51-hit.json", "w").write(json.dumps(
    body("Reply with exactly: CACHE-HIT-OK", 32)
))
print("wrote /tmp/fable51-warm.json and /tmp/fable51-hit.json")
PY

# --- bash ---
echo "=== WARM (expect cache_creation_input_tokens > 0, cache_read_input_tokens = 0) ==="
curl -sS https://api.anthropic.com/v1/messages \
  -H "content-type: application/json" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  --data-binary @/tmp/fable51-warm.json \
  | tee /tmp/fable51-warm-out.json | jq '{stop_reason, content, usage}'

echo "=== HIT (expect cache_read_input_tokens > 0) ==="
curl -sS https://api.anthropic.com/v1/messages \
  -H "content-type: application/json" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  --data-binary @/tmp/fable51-hit.json \
  | tee /tmp/fable51-hit-out.json | jq '{model, stop_reason, content: [.content[]?.text], usage}'

# --- bash ---
python3 - <<'PY'
import json
from pathlib import Path

def load(p):
    u = json.loads(Path(p).read_text())["usage"]
    return (
        u.get("cache_creation_input_tokens") or 0,
        u.get("cache_read_input_tokens") or 0,
        u.get("input_tokens") or 0,
        u.get("output_tokens") or 0,
    )

cw, cr, inn, out = load("/tmp/fable51-hit-out.json")
print(f"hit usage: cache_write={cw} cache_read={cr} uncached_input={inn} output={out}")
if cr == 0 and cw == 0:
    raise SystemExit("prefix was not cached; check that the system prompt is >= 512 tokens and the two bodies share the same system text")

def usd(tok, per_mtok):
    return tok / 1_000_000 * per_mtok

read_51 = usd(cr, 0.25)
read_5 = usd(cr, 1.00)
print(f"cache-read cost this turn: Fable 5.1 ${read_51:.6f} vs Fable 5 ${read_5:.6f} (saved ${read_5-read_51:.6f})")
print(f"plus uncached input ${usd(inn, 10):.6f} and output ${usd(out, 50):.6f} at 5.1 list")
PY

# --- bash ---
npm install -g @anthropic-ai/claude-code@2.1.258
claude --version   # expect 2.1.258

# one-shot session on the new default Fable id
claude --model claude-fable-5-1 -p --max-turns 1 --output-format json \
  "Reply with exactly: OK" | jq '{result, modelUsage}'

# gateway users: `fable` still means Fable 5 until the gateway is upgraded.
# pick 5.1 explicitly:
#   claude --model claude-fable-5-1
# or inside a session: /model claude-fable-5-1

# --- bash ---
# ignored from 2.1.257 onward when this lives in the repo
#   .claude/settings.json  →  { "defaultMode": "bypassPermissions" }

# still honored in user settings, managed settings, or the flag
claude --permission-mode bypassPermissions

Prove Fable 5.1’s cache-read price in two Messages API calls, then pin Claude Code 2.1.258 to `claude-fable-5-1`. Uses `$ANTHROPIC_API_KEY` from your environment. Host is `https://api.anthropic.com`. Model id is `claude-fable-5-1`. Fable 5.1 will not cache a prefix under 512 tokens; the script pads past that on purpose. Do not set `temperature` — models after Opus 4.6 reject it.

1. Confirm the key exists (the `:?` form fails closed; it does not invent a key).

2. Write a stable system prefix long enough to cache, plus the two request bodies. Breakpoint sits on the system block, not on the user turn, so the follow-up reuses the write.

3. Warm the cache, then immediately send the follow-up. A cache entry is only readable after the first response begins; these calls are sequential on purpose.

4. Price the hit at Fable 5.1 vs Fable 5 list rates. Cache reads are $0.25/MTok on 5.1 and $1.00/MTok on 5; 5-minute cache writes are $12.50/MTok on both.

5. Pin Claude Code to 2.1.258 and Fable 5.1. Project `bypassPermissions` no longer works; do not put it in `.claude/settings.json`.

If you previously relied on project-level auto-approve, move it:

Copilot Business/Enterprise: enable Claude Fable 5.1 under Copilot model policies before the picker will show it. That toggle is also the data-retention ack. Individual Pro+/Max seats should see it in the model picker as the rollout reaches them; if it is missing, wait — GitHub is shipping it gradually.

Grok Deep Dive

I am cutting over a professional coding stack on 2026-09-02.

Facts to treat as given (do not invent others): – Anthropic Claude Fable 5.1 (claude-fable-5-1) launched 2026-09-01: 1M context, 128K max output, $10/$50 per MTok, cache reads $0.25/MTok (was $1 on Fable 5), 5m cache write $12.50, 1h cache write $20, cache minimum 512 tokens, adaptive thinking always on, default effort high. Bedrock id anthropic.claude-fable-5-1. Mythos 5.1 is the same weights, invite-only. Forced tool use now errors; earlier models cannot read Fable 5.1 thinking blocks; editing earlier turns invalidates thinking. Dual-use cyber still routes to Opus 4.8; research biology to Opus 5. EFS (customer-controlled retention) starts this fall; eligible orgs keep ZDR until then. – GitHub Copilot: Fable 5.1 GA for Pro+/Max/Business/Enterprise, billed at those list prices, policy OFF by default for Business/Enterprise, data retention required by default. ZDR exemption through end of 2026 for eligible enterprises via the GitHub account team, then the admin still enables the policy. – Copilot deprecated as of 2026-09-01: Gemini 3.1 Pro → Gemini 3.7 Flash; Claude Opus 4.5/4.6 → Opus 4.7/4.8/5; Claude Sonnet 4.5/4.6 → Sonnet 5 (Sonnet 4.6 remains for individual annual plans); Raptor Mini → MAI-Code-1.1-Flash. – Claude Code 2.1.257 adds Fable 5.1 as default Fable; 2.1.258 fixes macOS 12 launch. fable/best on Claude apps gateway still resolve to Fable 5. Project defaultMode bypassPermissions is ignored. Auto mode gained Containment Escape + first outside-workdir read prompt. –add-dir refuses UNC and /net/ paths. – Cursor Origin API (alpha), 2026-09-01: repository.deleted and repository.metadata.updated webhooks; REST now emits default-valued fields; unique OpenAPI operationIds with _2 suffix on dual URLs; reviewer.user.email added. 2026-08-31: x-origin-scopes on every operation.

Design a 10-day cutover for a 40-engineer team that today pins Copilot to Claude Sonnet 4.6, runs Claude Code auto mode in CI with project-level bypassPermissions, and has a Cursor Origin GitHub App that treats missing JSON keys as false/empty. Cover: which Copilot model policy to enable (and the retention ack), which Copilot ids to rewrite before selectors go empty, whether to put Fable 5.1 on Copilot coding agent vs Claude Code vs the Messages API, a prompt-cache budget using the 0.025× read multiplier, the Claude Code settings move, and Origin client regeneration. Give a rollback that does not leave CI in bypassPermissions or Copilot on a retired id.