Claude Opus 5 launched yesterday. The benchmark comparisons are already everywhere — it matches or beats Fable 5 on agentic coding at half the price, as covered in our earlier post. What those posts skip is what you actually need to change in your codebase. Three new behaviors ship with the Claude Opus 5 API: an effort toggle with five levels, a Fast Mode that reaches 2.5x speed at a significant price premium, and an automatic fallback router that keeps your traffic moving when safety classifiers trigger. Here is what each one does and how to use it.
The Effort Toggle: Intelligence on Demand, Per Request #
The most significant API change in Opus 5 is the effort
parameter, passed inside output_config
. It controls reasoning depth — and, by extension, cost and latency — at the individual request level rather than the model level. Five options: low
, medium
, high
, xhigh
, and max
. The API defaults to high
. Full details are in the Anthropic effort level documentation.
This is not simply a “thinking dial.” The effort level governs all token spend: visible output text, extended thinking tokens, and the number of tool calls in an agentic loop. Lower effort produces fewer, more consolidated tool calls and terser responses. Higher effort means more planning steps, more tool use, and longer output. Each step up approximately doubles the cost of a call; going from low
to max
can be 10 to 20 times more expensive on a complex task.
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-opus-5",
max_tokens=64000,
messages=[{"role": "user", "content": "Refactor this module..."}]
)
response = client.messages.create(
model="claude-opus-5",
max_tokens=16000,
output_config={"effort": "medium"},
messages=[{"role": "user", "content": "Classify this support ticket..."}]
)
The practical rule: start at high
and evaluate from there. Step down to medium
for pipelines you have already validated — classification, routing, or structured extraction tasks where quality holds at lower effort. Use xhigh
only when your evals show a measurable gain over high
. Reserve max
for tasks where that last margin of accuracy justifies 45-60 second response times. low
is for latency-critical, scope-limited work like quick lookups or intent detection.
One practical note: at high
and above, set a generous max_tokens
value. It is a hard cap on thinking tokens plus response text combined. If you cap it too low, the model will truncate mid-thought.
Remove Your Verification Prompts #
Opus 5 verifies its own work without being asked. If your system prompt carries instructions like “double-check your answer,” “include a final verification step,” or “use a subagent to verify,” you are now paying for that verification twice — the model does it, then your instruction triggers it again. Those phrases should come out of your prompts.
Related: responses and written deliverables run longer on Opus 5 than on previous models. If your output token budget was sized for Opus 4.8 or Fable 5 behavior, revisit it. Underestimating max_tokens
here costs you quality, not just money.
Fast Mode: When 2.5x Speed Is Worth 6x the Price #
Fast Mode delivers up to 2.5x output tokens per second. The cost is $30 per million input tokens and $150 per million output tokens — six times the standard rate. It runs as a research preview on the Anthropic API and Claude Managed Agents only; Bedrock and Vertex AI do not have it yet. The Fast Mode documentation covers rate limits and current availability.
response = client.messages.create(
model="claude-opus-5",
max_tokens=8000,
extra_headers={"anthropic-beta": "fast-mode-2026-02-01"},
extra_body={"speed": "fast"},
messages=[{"role": "user", "content": "Complete this function..."}]
)
Fast Mode makes sense for real-time coding assistants, interactive agent loops where a human is waiting, or streaming completions where time-to-first-token is your constraint. It does not make sense for batch processing, background agents, or any pipeline where you can afford to wait. When the fast mode rate limit hits, requests automatically fall back to standard speed — your code does not need to handle that case explicitly.
Migrating from Fable 5 #
The model string swap is straightforward: replace claude-fable-5
with claude-opus-5
. Tool use schema is identical. Prompt structure carries over. The context window is the same 1M tokens. What actually breaks is behavioral, not structural. The Opus 5 official changelog documents the full list of behavioral changes.
Two things will catch you off guard. First, any prompt with explicit verification instructions will over-verify, as described above. Second, output responses trend longer — if you have downstream systems that parse or truncate model output by length, test them before switching. On the upside, safety refusals drop roughly 85% compared to Fable 5. The new automatic fallback router handles the cases that do trigger: instead of a content_policy_violation
error that kills your request, the API reroutes to an available model automatically.
For teams deploying on AWS, Claude Opus 5 is now available on Amazon Bedrock — though Fast Mode and the fallback router are Anthropic API-only features for now.
Cache-Safe Tool Switching #
A useful fix that shipped quietly: swapping tool sets mid-conversation no longer invalidates the prompt cache. Previously, adding or removing a tool from the tools array wiped the entire cache for that conversation. Multi-phase agentic pipelines — where a search phase and a coding phase use different tool sets — would lose cache benefits on every phase transition. That cost is gone. However, model switches still invalidate cache (caches are model-scoped), but tool-set changes are now safe.
Which Model to Use #
The honest answer: Opus 5 at high
effort is now the sensible default for complex agentic coding and computer use tasks. It beats Fable 5 on agentic benchmarks and costs half as much. Fable 5 retains a slim edge on SWE-bench Pro (80% vs 79.2%) and specific legal-agent workflows — if your evals show a meaningful gap there, stay on Fable 5. For cost-sensitive applications, Sonnet 5 at $3/$15 per MTok (introductory $2/$10 through August 31) is the right call; Opus-class capability is unnecessary overhead for most classification, extraction, and routing tasks.
The effort toggle signals something broader. Per-request intelligence granularity — rather than per-model selection — is where the Anthropic API is headed. The model hierarchy is compressing. The control surface is shifting from which model you pick to how much reasoning you buy per call.