Table of Contents
When people argue about the single most capable model you can call today, Claude Opus 4.8 is usually in the first sentence. Anthropic’s flagship, released in May 2026, currently ranks #1 on both the Artificial Analysis coding and intelligence indexes, pairs a 1-million-token context with vision and tool use, and is built for the autonomous, long-horizon work where coherence over a long session actually matters.
If you want to build on it, you can go direct to Anthropic, or reach it through a model aggregator such as OrcaRouter, which exposes Claude Opus 4.8 through a standard OpenAI-compatible endpoint. Either way, here’s the rundown.
TL;DR — Anthropic’s flagship. 1M context, 128K output, text + image + file input, reasoning + tools. $5 / $25 per million tokens. Ranks #1 on the Artificial Analysis coding (69.9) and intelligence (73.9) indexes — the model to beat for the hardest work.
Claude Opus 4.8 at a glance
- Developer: Anthropic — model id anthropic/claude-opus-4.8
- Released: May 28, 2026
- Context: 1,000,000 tokens; up to 128K output
- Inputs: text, image and file; reasoning, tools and JSON supported
- Price: $5 / 1M input, $25 / 1M output (prompt caching and batch discounts available)
- Built for: autonomous agents, long-horizon and knowledge work, computer use
Why Claude Opus 4.8 stands out
The headline is raw capability. Opus 4.8 is designed for the tasks that break lesser models: multi-hour agent runs, deep reasoning, computer-use workflows, and knowledge work where the model has to stay coherent across a very long session.
It’s also genuinely multimodal — text, image and file inputs — with a 1-million-token context, so it can read a whole repository, a stack of PDFs, or a design plus its spec and reason over all of it at once.
For teams already on Anthropic, prompt caching (a steep discount on repeated context) and batch processing keep the flagship’s cost manageable on high-volume jobs.
Benchmarks: the numbers behind the reputation
Opus 4.8 doesn’t just place well — it currently tops the charts. On Artificial Analysis it holds #1 of 110 on the Intelligence index (73.9) and #1 of 106 on Coding (69.9), better than 99% of models tracked. Supporting scores are strong across the board: GPQA Diamond 61.9, MMLU-Pro 72.9, TerminalBench Hard 58.3.
These are vendor/aggregator-reported, as always, so validate on your own tasks — but for the hardest reasoning and coding, this is the current high-water mark.
The short version: when the task is genuinely hard and getting it right matters more than the per-token price, Opus 4.8 is the model to reach for.
Pricing
At $5 / $25 per million tokens, Opus 4.8 is a premium model — but it’s priced below the tier above it, and its per-token output rate undercuts some rivals.
| Model | Input / 1M | Output / 1M | Context |
| Claude Opus 4.8 | $5.00 | $25.00 | 1M |
| GPT-5.5 | $5.00 | $30.00 | 1M |
| Claude Fable 5 | $10.00 | $50.00 | 1M |
The practical way to keep the bill sane is to reserve Opus 4.8 for the hard calls, lean on prompt caching for repeated context, and route routine traffic to a cheaper model — all from the same endpoint. Pay-per-token API access also means no daily subscription cap to ration against.
How to access the API
If your code already speaks the OpenAI format, switching is a two-line change, not a rewrite — point the base URL at the provider and set the model string:
| from openai import OpenAI
client = OpenAI( base_url=”https://api.orcarouter.ai/v1″, api_key=”$ORCAROUTER_API_KEY”, ) response = client.chat.completions.create( model=”anthropic/claude-opus-4.8″, messages=[{“role”: “user”, “content”: “Hello”}], ) print(response.choices[0].message.content) |
Because the endpoint is OpenAI-compatible, you can call Opus 4.8 from the OpenAI SDK, LangChain, Cursor or Cline, and keep a cheaper model configured for everyday work — switching is one model-string change.
Getting the most out of Claude Opus 4.8
A capable model rewards a bit of discipline. Be explicit in your prompts about format, constraints and what a good answer looks like — precise instructions beat vague ones and cut the expensive retries that quietly inflate a bill. And match effort and context to the job: a tighter, well-scoped prompt is usually faster and cheaper than dumping everything into a giant context window just because you can.
It also pays to think in tiers instead of using one model for everything. Route the routine bulk of your traffic to a cost-effective option and reserve Claude Opus 4.8 for the requests that genuinely need it. Because the endpoint is OpenAI-compatible, that routing is a configuration choice rather than a rewrite, so you can tune the cost/quality balance without touching application code.
Finally, keep a small evaluation set of your own real prompts and re-run it whenever you weigh a new model or setting. Public benchmarks are a useful starting point, but the only scores that matter are the ones on your own tasks — and testing a challenger is a one-line model-string change, so there’s little reason not to measure before you commit.
Who should use it — and who shouldn’t
Reach for it when: the task is the hardest kind — long-horizon agents, deep reasoning, multimodal analysis, computer use — and correctness beats cost. This is where the flagship earns its rate.
Look elsewhere when: you’re doing high-volume, routine work where a mid-tier model delivers most of the quality for a fraction of the price. Use Opus 4.8 as the escalation path, not the daily driver.
FAQ
Why does Opus 4.8 burn through my limits so fast?
It’s the #1 community complaint: Opus defaults to high-effort adaptive thinking, over-explains, and tends to rewrite an entire file for a two-line change. The single most effective fix people report is to instruct it to output only the modified code block and skip the explanations — that alone slashes output tokens on coding work.
How do I cut Opus 4.8 token costs?
Beyond the output-only-diff trick, the consensus playbook is: don’t use Opus for everything (route simple tasks to a cheaper model), keep a claude.md that defines your desired output style, manage your context aggressively, and use prompt caching for repeated context. One myth to skip — ‘launch subagents’ prompts are widely considered a placebo, not a real token saver.
Does Opus 4.8 feel like it overthinks or runs slow now?
For simple tasks, many users say yes — the adaptive thinking is overkill and it can feel slow and verbose. The fix is to match reasoning effort to the task and reserve Opus for genuinely hard work, sending the routine bulk to a mid-tier model.
How do I use Opus 4.8 without hitting Max-plan limits?
Call it through the API and pay per token instead of rationing against a subscription cap. On an OpenAI-compatible endpoint you set the model to anthropic/claude-opus-4.8, with no daily ceiling to burn through.