Anthropic released Claude Opus 5 on July 24, and it immediately shot to the top of Hacker News. The reason is a single number: 30.2% on ARC-AGI-3 — more than three times the previous best score by any model. It’s available now as claude-opus-5
via the API at $5/$25 per million input/output tokens, the same price as Opus 4.8 and exactly half of Fable 5.
The ARC-AGI-3 Score That Changes the Conversation #
ARC-AGI-3 is not a benchmark you can game. François Chollet designed it specifically to defeat memorization: models are dropped into interactive environments with no instructions, no rules, no stated goals, and must infer underlying patterns from a handful of examples. Prior to July 24, GPT-5.6 Sol held the record at 7.8%. Opus 4.8 was sitting at 1.5%.
Opus 5 scored 30.2%. That’s four times the previous best and twenty times its predecessor. When the benchmark is explicitly designed to measure genuine reasoning rather than pattern recall, a jump this size is hard to attribute to dataset overlap. The coding benchmarks reinforce the point: according to Anthropic’s launch announcement, Opus 5 hits 43.3% on FrontierBench v0.1, beating both Fable 5 (33.7%) and GPT-5.6 Sol (34.4%) on agentic terminal coding. SWE-bench Verified came in at 96.0%; SWE-bench Pro at 79.2%.
Anthropic describes the model as behaving “like a careful scientist” — building custom computer vision pipelines when faced with incomplete inputs, catching edge cases in open-source packages that competitors missed, achieving gold-medal performance on IMO 2026 math problems without any external tools. Whether these descriptions survive contact with your specific workload is something you’ll need to test. However, the benchmarks are consistent enough across independent sources that the performance narrative holds.
The Cost Story — With One Caveat #
The pricing headline is clean: $5 per million input tokens, $25 per million output tokens. That matches Opus 4.8 exactly and undercuts Fable 5 by 50%. For teams currently paying Fable 5 rates for tasks where Opus 5 reaches parity — and on FrontierBench it doesn’t just reach parity, it surpasses it — the economic case is straightforward.
However, “same per-token price” is not the same as “same total cost.” Historical pattern with the Claude lineup: Opus 4.7 consumed 30-40% more tokens per task than Opus 4.6 despite identical rates, because more capable models generate more thorough outputs. The effort toggle is Anthropic’s answer to this. You can dial response depth from low
through max
. Anthropic recommends low
and medium
for general tasks, xhigh
specifically for coding. Notably, max
can degrade performance in some cases while costing more. Measure your real task costs before counting the savings.
Related:[LLM API Costs Dropped 94%: What to Fix in Your Architecture Now]
What to Fix in Your Claude API Code Before Migrating #
Three changes in Opus 5 will break existing Opus 4.8 integrations if you switch without reading the migration docs. The biggest is thinking behavior: thinking is now enabled by default. On Opus 4.8, you had to opt in; on Opus 5, it activates automatically with depth controlled by the effort parameter. More critically, setting thinking: disabled
with effort xhigh
or max
now returns a 400 error. Code that relied on disabling thinking at high effort levels will fail in production until you catch this.
The fix is a two-option choice. Either keep thinking disabled and drop effort to high
or below, or remove the thinking configuration entirely and let the default run:
client.messages.create(
model="claude-opus-5",
thinking={"type": "disabled"},
extra_body={"effort": "xhigh"}
)
client.messages.create(
model="claude-opus-5",
thinking={"type": "disabled"},
extra_body={"effort": "high"}
)
client.messages.create(
model="claude-opus-5",
extra_body={"effort": "xhigh"}
)
Third change: delete your verification prompts. Instructions like “include a final verification step” or “check your work before responding” now cause over-verification. Opus 5 already verifies its work internally; stacking manual instructions on top produces bloated, redundant outputs. Anthropic’s prompting guide for Opus 5 explicitly advises removing these from existing system messages.
Related:[Claude Managed Agents Memory API Changed Today: Fix These 3 Things Now]
New Beta Features Worth Testing #
Two developer features shipped alongside the model, both in beta. Mid-conversation tool changes let you add or remove tools between turns without invalidating the prompt cache — a meaningful improvement for agentic applications that previously required fixed tool lists for an entire session. Use the beta header mid-conversation-tool-changes-2026-07-01
. Additionally, automatic fallbacks route safety-flagged requests to alternative models instead of returning errors; enable with server-side-fallback-2026-07-01
. Both are useful in production contexts, but test before deploying — beta status means APIs may shift.
Key Takeaways #
- Claude Opus 5 achieved 30.2% on ARC-AGI-3 on July 24 — tripling the previous best by any model and 20x higher than Opus 4.8, on a benchmark that cannot be gamed by memorization
- On agentic coding (FrontierBench v0.1), Opus 5 beats both Fable 5 and GPT-5.6 Sol at half the per-token price — the strongest case yet to stop paying Fable 5 rates for most coding workloads
- The $5/$25 per-million-token price matches Opus 4.8, but expect higher token consumption per task — measure real costs using the effort toggle before committing budgets
- Migrating from Opus 4.8 requires three fixes: handle the 400 error on xhigh+disabled-thinking, expect thinking to be on by default, and remove manual verification prompts from system messages