Model Lineup & Pricing

Current lineup (September 2026)

Cache writes cost ≈ 1.25× base input (5-minute TTL) or 2× (1-hour TTL). Batch API: 50% off input and output.

Legacy, still available: Fable 5, Opus 4.8, 4.7, 4.6, 4.5, Sonnet 4.6, 4.5. Retired: Opus 4.1 (2026-08-05).

Published retirement floors: Fable 5.1 not before 2027-09-01 · Opus 5 not before 2027-07-24 · Sonnet 5 not before 2027-06-30 · Haiku 4.5 not before 2026-10-15.

Capability matrix

Fable 5.1 breaking changes (exam favourites)

1. No forced tool use. tool_choice: "any" and {"type": "tool", "name": …} return 400. Alternatives: auto plus an explicit instruction; strict: true on the tool schema; output_config.format structured output.
2. Thinking-block binding. Thinking blocks are readable only by the model that produced them or a newer one. A fallback or router switch to an older model silently drops them (unbilled) and the older model re-plans from scratch.
3. Append-only history. Editing, reordering or removing earlier turns invalidates every later thinking block. Freeze system and tools, deliver mid-session instruction changes as role: "system" messages, and trim context server-side (context editing / compaction) rather than on the client.
4. Retention. Requires 30-day data retention; not available to ZDR organisations unless authorised. Excluded from Priority Tier (as are Opus 5 and Sonnet 5).

Selecting a model – decision table

Worked cost calculations

Example 1 – 10,000 documents, 3,000 input tokens each, 500 output tokens each

Example 2 – prompt caching on a 20,000-token system prompt + tools, 1,000 requests/day on Sonnet 5

Rule: caching pays back after roughly two reads of the same prefix within the TTL.

Example 3 – choosing effort

Cloud availability

Feature availability can lag on cloud platforms; check per-feature docs before committing an architecture.

Feature availability by cloud platform

The exam tests the reflex that the direct Claude API gets new features first, and that a governance or residency requirement can force a platform that lags on a feature you depend on. Treat this as a snapshot to reason with, not a live SLA.

Rate-limit tiers and dimensions

Rate limits apply on three dimensions simultaneously; you hit whichever binds first.

Deprecation and migration lifecycle

Model retirement is a planned lifecycle event, not an emergency. The exam-correct posture: watch deprecation notices, keep an eval set, migrate behind that eval, pin IDs in production.

Announced → Deprecated (still callable) → Retirement floor date → Retired (404)
   │              │                              │
   │              └─ start eval-gated migration  └─ published earliest date; often extended
   └─ appears in models.list() deprecation metadata / dashboard

Per-model migration checklists

1. To Haiku 4.5 — remove any effort parameter (unsupported → 400); if you relied on budget_tokens, this is the only current model that keeps it, so a downgrade from adaptive-thinking models is where budget_tokens reappears. Re-tune prompts for the 200k/64k window; verify classification/extraction accuracy on the golden set.
2. To Sonnet 5 — drop any mid-conversation role: "system" messages and task budgets (unsupported); confirm xhigh effort is not assumed (no benefit). Re-check latency budgets; it is the balanced default.
3. To Opus 5 — safe target for most agentic upgrades; xhigh available. Remove budget_tokens (400) in favour of {"type":"adaptive"}. Re-baseline cost — 2.5× Sonnet input.
4. To Fable 5.1 — the highest-friction migration. Remove forced tool_choice (any/named → 400); switch to auto+instruction, strict: true, or output_config.format. Freeze system/tools for append-only history. Confirm the org is not ZDR (30-day retention required) and does not depend on Priority Tier. Re-baseline cost at $10/$50.

More worked cost scenarios

Example 4 – caching + Batch combined

Nightly enrichment of 50,000 records on Sonnet 5. Each request: 12,000-token shared instruction/schema prefix (cacheable) + 800 unique input tokens + 300 output tokens. Run as one Batch job.

Naïve Sonnet 5 realtime, no cache: prefix 50k×12k=600M×$2=$1,200 + input 40M×$2=$80 + output 15M×$10=$150 = $1,430. Caching+Batch cuts it ≈ 88%. The prefix dominates, so caching it matters far more than the discount on the small unique portion.

Example 5 – cascade routing math

A stream of 100,000 tickets. A Haiku 4.5 first pass answers 100% (1,500 in / 400 out each). An external validator flags 18% as low-confidence; those escalate to Opus 5 (same tokens). Compare to sending everything to Opus 5.

Cascade saves ≈ 62%. The break-even escalation rate e where cascade = Opus-only solves 350 + 1750e = 1750 → e ≈ 80%. Below ~80% escalation, cascade wins; above it, the Haiku pass is pure overhead — send everything to Opus.

Common misconceptions

Scenario walkthrough

A fintech runs a document-classification pipeline on 2M PDFs/month. Requirements: US-federal customer (FedRAMP High), personal data (needs regional residency and, ideally, ZDR), cost-sensitive, latency non-critical (nightly), classification quality must be measured per document type before switching models.

Expert reasoning trace:

1. Residency + FedRAMP High rules out the direct API for the regulated workload → Bedrock or Vertex. Pick the one matching existing cloud spend/IAM.
2. ZDR desired + cheap + high-volume classification → Haiku 4.5. It is Priority-Tier eligible (irrelevant here, latency non-critical) and supports ZDR. Fable 5.1 is rejected: no ZDR, 10× the price, and classification does not need frontier reasoning (constraint-blind + over-engineered).
3. Latency non-critical, 2M/month → Batch API (50% off). Realtime is rejected as needlessly expensive (cost-blind).
4. Shared schema/instruction prefix → prompt caching on the stable prefix; dynamic PDF last.
5. “Measured per document type” → per-segment eval on a golden set, not aggregate accuracy (aggregate-metric distractor).
6. Model migration later → pin claude-haiku-4-5, keep the eval set, migrate behind it when a successor ships.

Correct architecture: Bedrock + Haiku 4.5 + Batch + prompt caching + per-segment evals. Each tempting alternative (Fable everywhere, realtime for speed, aggregate accuracy, self-reported confidence routing) maps to a named distractor pattern.