Claude Opus 5 (claude-opus-5) shipped July 24, 2026 at $5/$25 — half Fable 5's price. Benchmarks, the Opus 4.8 to 5 API changes, and how to call it via ofox.
What Anthropic Shipped
Claude Opus 5 is Anthropic's new recommended default for complex agentic coding and enterprise work. The model ID is claude-opus-5 with a fixed, dateless snapshot and no version suffix. It carries a full 1M-token context window, 128K maximum output tokens (expandable to 300K through Message Batches API with beta header), and May 2026 knowledge cutoff.
The pricing represents a significant capability-to-cost ratio: "a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price." Fable 5 costs $10/$50 per million tokens; Opus 5 lists at $5 input/$25 output—matching Opus 4.8's rate since May and undercutting Fable 5 by fifty percent. Fast Mode runs the same model at up to 2.5x output speed for double the base cost.
Fable 5 remains Anthropic's most capable model. Opus 5 targets users seeking most of that capability without the Fable premium.
Benchmarks: Anthropic's Own Numbers
All figures presented are Anthropic internal and unverified vendor launch numbers, not third-party results. Treat them as best-case scenarios and run independent evaluations before commitment.
| Benchmark | Anthropic internal figure (unverified) |
|---|---|
| Frontier-Bench v0.1 | Surpasses all other models; more than doubles Opus 4.8's score |
| CursorBench 3.2 | Within 0.5% of Fable 5's peak, at half the cost |
| ARC-AGI 3 | Roughly 3x the next-best model |
| OSWorld 2.0 (computer use) | Beats Fable 5's result at one-third the cost |
| Zapier AutomationBench | Pass rate ~1.5x the next-best model |
| GDPval-AA (real economic work) | State of the art |
The consistent theme spans agentic and computer-use work—long-horizon coding, desktop control, real-task automation—where Opus 5 either matches Fable 5 for substantially less money or outperforms competing models. Anthropic acknowledges one honest limitation: Opus 5 stays behind Mythos 5 on cybersecurity-exploitation and biology-research tasks, handled by a separate model line for defensive use. Non-applicable workloads face no gap.
What Changed from Opus 4.8 (the part that breaks code)
Opus 5 maintains the same request surface as 4.8 in most respects, but three changes will surprise users swapping the model string without additional modifications.
1. Adaptive thinking is on by default
On Opus 4.8, a request without a thinking field ran without thinking. On Opus 5, the identical request now runs with adaptive thinking. Since max_tokens is a hard cap on total output—thinking plus visible response—workloads tuned for no-thinking 4.8 can truncate. Revisit max_tokens, or pass thinking: {"type": "disabled"} to retain the old behavior.
2. Disabling thinking is capped at high effort
This is the new 400 error users don't expect. You can disable thinking, but only at effort high or below. Combining thinking: {"type": "disabled"} with effort xhigh or max triggers rejection:
# Rejected on Opus 5 — 400 error
client.messages.create(
model="claude-opus-5",
max_tokens=16000,
thinking={"type": "disabled"},
output_config={"effort": "xhigh"}, # xhigh/max + disabled = 400
messages=[{"role": "user", "content": "..."}],
)
# Fix: either re-enable thinking at high effort...
output_config={"effort": "xhigh"} # thinking on by default
# ...or keep it off and drop to high or below
thinking={"type": "disabled"}, output_config={"effort": "high"}
3. The cache minimum drops to 512 tokens
Cache thresholds decreased from 1,024 on 4.8. Short system prompts previously too small to cache now create cache entries without code modification—a small, free win on prompt-caching expenses.
Two additional considerations: mid-conversation tool changes are now possible without invalidating the prompt cache (beta header mid-conversation-tool-changes-2026-07-01), and Opus 5 ships cybersecurity safety classifiers, requiring handling of stop_reason: "refusal" with optional automatic model fallbacks (fallbacks: "default", beta header server-side-fallback-2026-07-01).
Unchanged elements: pricing, the 1M context window, 128K output, and hard rejections from 4.8—non-default temperature/top_p/top_k, manual budget_tokens, and assistant prefill all return 400. Priority Tier remains unsupported.
Effort and Prompting Notes
Opus 5 supports the complete effort ladder—low, medium, high, xhigh, max—defaulting to high on Claude API and Claude Code. Anthropic recommends running a fresh effort sweep rather than carrying over 4.8 settings, and testing max for capability-critical work (allocate large max_tokens, starting around 64K, for xhigh or max).
Two behavioral shifts warrant prompt review. Opus 5 verifies its own work without explicit instruction, so explicit "double-check your output" directives now cause over-verification—remove them. Default responses run longer than 4.8's; lowering effort trims thinking but not visible length, so prompt for conciseness or target length directly for shorter output.
How to Call Opus 5 Through ofox
The model ID on ofox.ai is anthropic/claude-opus-5, on the same OpenAI-compatible endpoint as every other model—no separate Anthropic account or billing. Point your existing OpenAI SDK at the ofox base URL and change two strings:
from openai import OpenAI
client = OpenAI(base_url="https://api.ofox.ai/v1", api_key="your-ofox-key")
response = client.chat.completions.create(
model="anthropic/claude-opus-5",
messages=[{"role": "user", "content": "Refactor this module for testability..."}],
)
To use adaptive thinking and the effort parameter, call the Anthropic-native endpoint instead—the model ID drops the provider prefix:
import anthropic
client = anthropic.Anthropic(
base_url="https://api.ofox.ai/anthropic",
api_key="your-ofox-key",
)
response = client.messages.create(
model="claude-opus-5",
max_tokens=8000,
thinking={"type": "adaptive"},
output_config={"effort": "high"}, # low | medium | high | xhigh | max
messages=[{"role": "user", "content": "Audit this service for race conditions..."}],
)
Running through one gateway makes the migration question empirical: put identical prompts through Opus 5, Opus 4.8, and Fable 5 on one key and compare quality and token counts on your workload before switching production.
Opus 5 vs Fable 5 vs Sonnet 5: Which to Pick
- Claude Opus 5 ($5/$25)—the new default for agentic coding, computer use, and long-horizon enterprise tasks. Best capability-per-dollar in the lineup right now.
- Claude Fable 5 ($10/$50)—the ceiling. Reach for it only when a task genuinely needs the most capable model and cost is secondary; on many agentic benchmarks Opus 5 is within a rounding error at half the price.
- Claude Sonnet 5 ($3/$15)—the speed-and-cost tier. Near-Opus quality on coding for high-volume or latency-sensitive work where you don't need Opus-level reasoning depth.
For Opus 4.8 users, this is a clean migration: same price, higher scores, one model-string change plus the three gotchas above.
Verdict
Opus 5 is another Opus upgrade with no asterisk on price—same $5/$25, materially higher scores across coding and computer-use, and Anthropic's claim of near-Fable-5 capability at half the cost holds up across a broad benchmark set. The caveats are narrow and honest: benchmark numbers are Anthropic internal and unverified (run your own evals), disabling thinking now conflicts with high effort levels, and default output runs longer so watch your max_tokens and latency budgets.
For new projects, start on Opus 5. For anything in production on 4.8, change the model string, revisit max_tokens, audit any request that disables thinking, and ship.
Originally published on ofox.ai/blog.
Top comments (0)