Iro AIIro AI Blog
Claude Opus 5: what actually changed, and who it's for
Anthropic's newest model lands at half the price of Fable 5 with a 1M-token context window. Here's what's genuinely new, what the benchmarks mean, and whether you should switch.
In this postThe short answerWhat's actually newWhat the benchmarks sayOpus 5 vs Fable 5 vs GPT-5.6Should you switch?
The short answer
Claude Opus 5 is Anthropic's cost-efficient frontier model, released on 24 July 2026 at $5 per million input tokens and $25 per million output tokens — half the input price of Fable 5, with capability Anthropic positions as approaching it in many categories. It ships with a 1M-token context window, 128K maximum output, and a May 2026 knowledge cutoff.
If you use Claude in a browser or an app, the practical change is that a stronger model is now the one answering you. If you build on the API, the change is a pricing decision: Opus 5 costs the same as Opus 4.8 while performing meaningfully better, which makes staying on 4.8 hard to justify.
Pricing and benchmark figures here reflect Anthropic's published numbers and reporting at the time of writing (26 July 2026).
What's actually new
The price is the product
Anthropic held Opus 5 at Opus 4.8's rate — $5/$25 per million tokens — while pushing capability toward Fable 5, whose input pricing sits at double that. Batch API pricing runs $2.50/$12.50 per million tokens for work that can tolerate delay. Read the release as a repositioning rather than a new ceiling: the frontier is being made affordable rather than pushed further out.
A 1M-token context window
One million tokens is roughly 750,000 words, which in practice means an entire codebase, a year of meeting notes, or several long reports in a single conversation without chunking them. Combined with 128K maximum output, it makes long-document work substantially less fiddly — you stop engineering around the context limit and start just pasting the material in.
Fast mode and effort settings
Two controls change how you spend money per request. Fast mode runs about 2.5x quicker at double the price ($10/$50 per million tokens) — worth it for interactive work where waiting is the bottleneck, wasteful for batch jobs. Effort settings let you dial reasoning depth up or down per request, so a simple reformatting task and a hard architectural question no longer cost the same.
A May 2026 knowledge cutoff
Opus 5 has the most recent training cutoff of any Claude model. That matters less than people assume — any model's knowledge of recent events is patchy near its cutoff, and anything genuinely current still needs a web search or a source you provide. But it does reduce the number of times the model confidently describes a superseded version of a fast-moving tool.
Practice this, don't just read it.
Iro AI turns ideas like the ones in this post into 5-minute exercises with feedback. Free tier, Pro from $1.15/week ($59.99 a year, 7-day free trial).
What the benchmarks say (and what they don't)
Anthropic reports Opus 5 more than doubling Opus 4.8's score on Frontier-Bench v0.1, landing within 0.5% of Fable 5's peak on CursorBench 3.2 at half the cost per task, and scoring roughly three times the next-best model on ARC-AGI 3. Anthropic also describes it as its most aligned model to date, with the lowest measured rates of deceptive behaviour.
Two honest caveats belong next to any of those numbers:
- Vendor benchmarks are marketing artifacts. Every lab picks the evaluations where its model looks strongest. That does not make the numbers false, but it does make cross-lab comparisons from press materials unreliable.
- Benchmark gains rarely map to your work. A model that scores three times higher on an abstract reasoning benchmark does not write emails three times better. For most everyday tasks — drafting, summarizing, explaining — the top handful of models are close enough that your prompt matters more than your model.
The number that is straightforwardly meaningful is price: $5/$25 versus Fable 5's premium tier is a real, verifiable difference that shows up on your bill every month.
Opus 5 vs Fable 5 vs GPT-5.6
| Model | Price (in/out per M) | Best for | Watch-out |
|---|---|---|---|
| Claude Opus 5 | $5 / $25 | Long-context work, everyday reasoning, best capability-per-dollar in Anthropic's line-up | Not the outright capability leader — Fable 5 still edges it |
| Claude Fable 5 | Premium tier | The hardest problems where cost is not the constraint | Roughly double Opus 5's input price; extra safety guardrails can fall back to another model on high-risk prompts |
| GPT-5.6 Sol | $5 / $30 | Long-horizon agentic work and frontier reasoning; Ultra mode for the hardest tasks | Higher output price than Opus 5; overkill for routine work |
| GPT-5.6 Terra | $2.50 / $15 | The direct cost-efficient competitor to Opus 5 for everyday tasks | Shorter context than Opus 5's 1M window |
The genuinely competitive matchup is Opus 5 against GPT-5.6 Terra and Sol: all three are aimed at the same buyer, who wants near-frontier quality without paying frontier prices. Opus 5's differentiator is the 1M-token context window; GPT-5.6's is the three-tier structure that lets you drop to Luna for cheap bulk work inside the same family.
Should you switch?
If you're on Claude Opus 4.8: yes, effectively automatically. Same price, better model, longer context. There is no argument for staying.
If you're on Fable 5 for everything: probably worth testing Opus 5 on your actual workload. If quality holds on the tasks you care about, you halve your input costs. Keep Fable 5 for the genuinely hard subset.
If you use ChatGPT and are wondering whether to move: almost certainly not worth the disruption on capability grounds alone. GPT-5.6 and Opus 5 are close enough that switching tools costs you more in lost muscle memory than you gain in benchmark points. Switch if you specifically need the 1M-token context, or if you prefer Claude's writing style — not because a chart moved.
If you're a casual user: nothing to do. You are already getting the newer model in the apps that use it, and the meaningful gap between you and someone getting more from AI has never been the model version. It's how well you prompt it, and whether you check the output.
Practice this, don't just read it.
Iro AI turns ideas like the ones in this post into 5-minute exercises with feedback. Free tier, Pro from $1.15/week ($59.99 a year, 7-day free trial).