Claude Opus 5.5 Launches Cheaper and Beats GPT-6 Astra

Anthropic shipped Claude Opus 5.5 on September 22, 2026, and led with a number most model launches bury: the price went down. Base API rates are now $4 per million input tokens and $20 per million output tokens, 20% below Opus 5, with cached reads cut 60% to $0.20 per million tokens. On Anthropic's own benchmark table, Opus 5.5 also beats OpenAI's flagship GPT-6 Astra on several agentic and knowledge-work tests, while costing roughly 40% less to match Astra where the two are close.

Claude Opus 5.5 pricing versus GPT-6 Astra

The number that matters for anyone footing an API bill: Opus 5.5 runs at $4/$20 per million input/output tokens, against GPT-6 Astra's $10/$50. Anthropic says Opus 5.5 at its new default effort level beats Opus 5 running at max effort for about a fifth of the cost, and matches GPT-6 Astra on Terminal-Bench 4.0 for around 40% of what Astra costs to hit the same score. Batch pricing drops further to $2/$10, and a faster inference tier runs $8/$40.

Anthropic Opus 5.5 benchmarks

Opus 5.5 leads Astra on Terminal-Bench 4.0 (66.4% vs 57.9%), FrontierCode v1.1 (54.4% vs 53.3%) and Humanity's Last Exam (67.7% vs 57.2%). Astra pulls ahead on AutomationBench (41.4% vs 40.0%) and Terminal-Bench-Science (64.6% vs 58.7%), which tracks with Astra's reported edge in agentic science and business-workflow automation. Anthropic itself cautions that "benchmark margins have become a less reliable guide" at this level of capability, and a four-to-two split across the seven benchmarks both vendors publish backs that up: this is a model that's clearly ahead on some tasks, not a clean sweep.

What it means for content and research pipelines

Two changes matter more than the leaderboard for anyone running AI on marketing content and research workflows. First, the 1 million token context window ships with no long-context surcharge, so pulling a full brand style guide, a quarter of campaign data and a competitor's site into one prompt no longer costs extra just for the length. Second, effort is now a first-class request parameter with five levels (low, medium, high, xhigh, max), and Opus 5.5 defaults to medium instead of the high default every earlier Opus model used. That means a request that simply omits effort now runs a notch leaner than it did on Opus 5, which is worth knowing before you assume nothing changed.

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=4096,
    messages=[
        {
            "role": "user",
            "content": "Summarize this quarter's campaign performance data "
                        "and flag any channel missing its CPA target.",
        }
    ],
    output_config={"effort": "high"},
)

for block in response.content:
    if block.type == "text":
        print(block.text)

Set output_config.effort to low for high-volume, on-brand social copy, and step it up to high or xhigh for a research brief that has to survive a client's fact-check. Thinking can't be disabled on Opus 5.5 (a request with thinking: {"type": "disabled"} returns a 400 error at every level), so effort is the only lever you get for cost and depth.

What to do Monday morning

If you're running content or research workflows on Opus 5 or Sonnet today, rerun your eval set against claude-opus-5-5 before touching a single prompt. The price drop alone changes the math on tasks you'd routed to a cheaper model for budget reasons, and the bigger context window means some multi-step retrieval pipelines might collapse into a single call. Don't carry your old effort settings over: Anthropic recommends a fresh effort sweep on Opus 5.5 rather than reusing what worked on Opus 5, since the default effort level itself has changed.

Cheaper frontier intelligence with a real context window bump is the kind of release that quietly resets what's economical to automate, and that's worth an afternoon of re-testing before your next sprint.