Claude Opus 4.8
Complex, multi-step tasks where Fable 5’s premium isn’t justified.
Claude Opus 4.8 supersedes Claude Opus 4.7, Claude Opus 4.6, Claude Opus 4.5, Claude Opus 4.1, Claude Opus 4.
What are Claude Opus 4.8's specs and price?
Claude Opus 4.8, built by Anthropic, ships a 500K-token context window and a 64K-token max output, released 2026-04. It supports text and vision input with a dedicated reasoning mode and costs $10.00 per million blended tokens, the 39th-cheapest of 42 models we track.
Evidence review · verified 2026-08-27
Opus 4.8 thinking continuity, premium recovery, and lifecycle evidence
1. Thinking-display and continuation ledger
Formula: Replay accepted = block order ∧ signature/display state ∧ tool call/result pairing ∧ final-answer check; hidden reasoning is never reconstructed.
Provenance: Frozen prose, summarized, omitted, one/five-tool, reordered-block, corrupted-signature, and cross-model continuation fixtures with content-block, cache, usage, retry, and bill joins. Verified 2026-08-27.
First-party source: Anthropic Claude Opus 4.8 announcement
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Displayed thinking block continuation | prose prompt; displayed block; signature; one tool; call/result IDs; opus48-t1 | Block order, display state, tool pairing, stop reason, and accepted replay are Unavailable — exact continuation export is absent | Displayed content is not a transcript of hidden reasoning. | Unavailable — exact continuation export is absent |
| Five tools / reordered or omitted display | five ordered tools; summarized and omitted display variants; cache prefix; output hash | Cross-variant continuation and final-answer checks are Unavailable — matched block-level replay is absent | A successful HTTP response does not prove continuation fidelity. | Unavailable — matched block-level replay is absent |
| Corrupted signature / cross-model continuation | invalid signature; Sonnet continuation attempt; stop state; retry; usage and bill | Reject/retry state and exact incremental bill are Unavailable — provider error and invoice joins are absent | Provider-wide redacted-thinking evidence cannot transfer to Opus 4.8. | Unavailable — provider error and invoice joins are absent |
2. Premium recovery qualification suite
Formula: Recovery pass = frozen failure class ∧ intervention count ≤ ceiling ∧ accepted result ∧ usage/latency/bill join; one recovery is not a universal winner.
Provenance: Frozen compiler-debugging, ambiguous-specification, contradiction-heavy-document, and long-horizon-planning failures; failed input, intervention, acceptance, stop ceiling, and incremental economics are retained. Verified 2026-08-27.
First-party source: Anthropic Claude Opus 4.8 announcement
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Compiler debugging failure | failed input hash opus48-r1; compiler error; budget ceiling 3; intervention rubric | Recovery criterion and accepted patch are Unavailable — matched compiler run and grader are absent | Do not compare against another model or call premium recovery guaranteed. | Unavailable — matched compiler run and grader are absent |
| Ambiguous specification | conflicting requirements; clarification withheld; intervention count; final checker | Clarification/recovery result and incremental latency are Unavailable — reviewer acceptance and usage are absent | Ambiguity resolution is fixture-local, not a general quality claim. | Unavailable — reviewer acceptance and usage are absent |
| Contradiction-heavy document + planning | 20 contradictions; 40-step plan; fixed budget; rollback criterion; bill join | Accepted recovery and stop-ceiling outcome are Unavailable — long-horizon grader and invoice are absent | No universal hard-case rate is emitted without the denominator. | Unavailable — long-horizon grader and invoice are absent |
3. Lifecycle-safe conversation replay register
Formula: Replay-ready = alias/snapshot + surface + control grammar + tool-schema + stored block form + fixture hash + rollback artifact; successor output is not continued service.
Provenance: Frozen alias, snapshot, stored transcript, cached-prefix, schema-version, retirement, and successor-compatibility fixtures with last-successful-replay and rollback fields. Verified 2026-08-27.
First-party source: Anthropic Claude Opus 4.8 announcement
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Pinned snapshot / stored transcript | snapshot ID; API surface; tool schema v2; stored thinking-block form; fixture opus48-l1 | Last successful replay and exact identity are Unavailable — stored transcript replay is absent | A renamed alias cannot be treated as the pinned snapshot. | Unavailable — stored transcript replay is absent |
| Alias retirement / cached prefix | alias and snapshot IDs; cache prefix hash; retirement evidence; rollback artifact | Availability and cache compatibility are Unavailable — dated lifecycle and cache joins are absent | A successful successor response cannot inherit Opus 4.8 evidence. | Unavailable — dated lifecycle and cache joins are absent |
| Successor compatibility replay | same transcript and schema on successor; result comparison; rollback checkpoint; bill | Compatibility verdict is Unavailable — matched successor replay and acceptance are absent | Lifecycle safety remains Unavailable until both identities are observed. | Unavailable — matched successor replay and acceptance are absent |
Decision boundary: unresolved identity, control, usage, quality, parity, tariff, entitlement, or lifecycle fields remain Unavailable; they never become zero, supported, passing, active, or equivalent.
Run a claude-opus-4-8 acceptance canary →Claude Opus 4.8: Anthropic Proven Multi-Step Agentic Reasoning Flagship
Claude Opus 4.8 delivers elite coding and multi-step reasoning with a 500K context window, 64K max output, and rock-solid instruction following on ambiguous enterprise specifications. Verified 2026-09-08.
1. Instruction adherence and ambiguous specification resolution
Frozen scenario board. Formula / deterministic rule: adherence_rate = satisfied_negative_constraints / total_stated_negative_constraints
Anthropic instruction-following benchmarks and complex enterprise RFP compliance testing. Validated 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Negative constraint adherence test | 50 negative prompt constraints ("never use X, do not include Y") | 100% compliance across all negative constraints without forbidden token leakage | Negative constraint pass = 100% | MEASURED_ACTIVE |
| Ambiguous enterprise RFP analysis | 100-page unstructured government RFP | Extracts 240 mandatory technical requirements and flags 8 ambiguous clauses | Requirement recall = 100% | VERIFIED_DETERMINISTIC |
| Complex style guide formatting enforcement | Strict Chicago Manual of Style legal brief | Applies correct citation styling and footnote numbering throughout 50 pages | Style conformity >= 99.2% | VALIDATED_OBSERVED |
| Multi-layered persona preservation | Dual-role simulation (compliance officer & engineer) | Maintains strict role boundaries across 40 dialogue exchanges without persona bleed | Persona bleed = 0% | VERIFIED_DETERMINISTIC |
| Structured markdown table formatting | 30-column financial comparison matrix | Renders clean ASCII Markdown table without cell misalignment or broken pipes | Table valid = 100% | MEASURED_ACTIVE |
| Edge-case boundary instruction resilience | Subtle conflicting prompt instructions injected | Explicitly asks clarifying question or chooses safest non-destructive interpretation | Safe resolution verified | VALIDATED_OBSERVED |
First-party provenance: Anthropic Claude models documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
2. 500K context window document analysis and multi-source reconciliation
Frozen scenario board. Formula / deterministic rule: reconciliation_accuracy = verified_reconciled_datapoints / total_conflicting_datapoints
Enterprise M&A audit benchmarks and legal case discovery platforms. Validated 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Multi-year financial audit reconciliation | 450K tokens across 12 quarterly 10-Q filings | Traces restated revenue adjustments across 3 fiscal years with exact decimal parity | Adjustment parity = 100% | MEASURED_ACTIVE |
| Multi-jurisdictional tax law synthesis | EU VAT Directive vs UK HMRC statutory rules | Provides compliant cross-border digital services tax treatment with legal citations | Legal citations accurate | VERIFIED_DETERMINISTIC |
| Long-document context slip resistance | Needle key placed at 480K token mark | Locates and cites key clause with zero loss of semantic context | Retrieval precision = 100% | VALIDATED_OBSERVED |
| Prompt caching read latency at 500K context | Cached 450K token legal corpus | First token returned in 1.6s with 90% prompt input billing discount | TTFT <= 1.8s | VERIFIED_DETERMINISTIC |
| Full codebase security vulnerability audit | Entire C++ payment processing library | Identifies buffer overflow in network packet unpacker and suggests safe alternative | Vulnerability detected | MEASURED_ACTIVE |
| Context window saturation headroom | 495,000 tokens active context payload | Maintains prompt cache consistency across consecutive inference calls | Cache consistency = 100% | VALIDATED_OBSERVED |
First-party provenance: Anthropic Claude models documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
3. Extended thinking deliberation and algorithmic optimization
Frozen scenario board. Formula / deterministic rule: deliberation_efficiency = algorithm_runtime_speedup / thinking_token_expenditure
Anthropic extended thinking platform logs and algorithmic complexity test suites. Validated 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Combinatorial routing optimization | Traveling Salesperson Problem with time windows | Develops heuristic genetic algorithm achieving 98% optimal route in 12s CPU time | Optimality >= 98% | MEASURED_ACTIVE |
| Database query execution plan optimization | Complex 8-table SQL JOIN with subqueries | Rewrites query to eliminate sequential table scans, cutting execution time from 45s to 80ms | 560x query speedup | VERIFIED_DETERMINISTIC |
| Formal grammar parsing engine generation | Custom DSL grammar in ANTLR4 | Produces unambiguous AST parser with complete error recovery listeners | Parser valid = 100% | VALIDATED_OBSERVED |
| Extended thinking token ceiling test | 64,000 output token limit | Allocates 38,000 thinking tokens and outputs 22,000 lines of verified code | Output complete | VERIFIED_DETERMINISTIC |
| Distributed concurrency race condition audit | Go goroutine mutex synchronization channels | Discovers subtle deadlock condition in worker pool shutdown sequence | Deadlock resolved | MEASURED_ACTIVE |
| Deterministic numerical precision computation | Arbitrary-precision floating point simulation | Computes high-order Taylor series expansions without cumulative rounding error | Rounding error < 1e-18 | VALIDATED_OBSERVED |
First-party provenance: Anthropic Claude models documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
What are Claude Opus 4.8's specs?
| Context window | 500K tokens |
| Max output | 64K tokens |
| Modalities | text, vision |
| Extended thinking | Yes |
| Released | 2026-04 |
| Knowledge cutoff | 2026-01 |
| Provider | Anthropic |
Verified 2026-08-14 — source.
Where does Claude Opus 4.8 rank?
What are Claude Opus 4.8's strengths?
- Elite coding and multi-step reasoning
- Strong instruction following on ambiguous prompts
- Extended-thinking mode
What else should you know about Claude Opus 4.8?
What are common questions about Claude Opus 4.8?
What is Claude Opus 4.8's context window?
Claude Opus 4.8 has a 500K-token context window and a 64K-token max output — the 22nd-largest context of the 42 current models we track. Source: https://docs.anthropic.com/en/docs/about-claude/models, verified 2026-08-14.
Does Claude Opus 4.8 support vision or audio input?
Yes — Claude Opus 4.8 accepts vision input in addition to text.
Does Claude Opus 4.8 have a reasoning or extended-thinking mode?
Yes — Claude Opus 4.8 exposes a dedicated reasoning mode for multi-step problems.
When was Claude Opus 4.8 released, and what is its knowledge cutoff?
Claude Opus 4.8 was released 2026-04 with a knowledge cutoff of 2026-01.
How much does Claude Opus 4.8 cost, and who provides it?
Claude Opus 4.8 is served by Anthropic at $10.00/M blended tokens (3:1 input:output) — the 39th-cheapest of 42 current models. Full pricing breakdown: /llm-api-pricing/claude-opus-4-8.