Claude Opus 5.5 shifts the AI race toward cheaper, longer work
Claude Opus 5.5
Anthropic's AI model for complex, long-running work.
token
A small unit of text used to measure work and price.
benchmark
A fixed test used to compare model performance.
What happened
Anthropic, the company that makes Claude, announced Claude Opus 5.5 on September 22, 2026. It is the first model in the Claude 5.5 family. Anthropic presents it as a model for long, difficult work. Its examples include coding, computer use, research, and business workflows. The company says it is available through the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic's official announcement
The launch also drew strong attention on Hacker News. The main submission reached 1,147 points and 784 comments. A second submission reached 273 points and two comments. These numbers measure community attention. They do not prove the model's quality or the announcement's accuracy. The main Hacker News submission
The background
Anthropic's announcement makes efficiency a central part of the story. AI systems can look impressive on one test. Long jobs also depend on steps, tokens, tools, and total cost. Anthropic says benchmark gaps are becoming less reliable at this capability level. It says the real difference between close models can be smaller than their scores suggest.
Why it matters
Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Opus 5 costs $5 and $25. Cache reads cost $0.20, down from $0.50. Anthropic says typical workloads cost 40% less. It also says output is more than 30% faster.
One early tester used Opus 5.5 on a 200,000-line codebase. The task took less than three hours. Opus 5 took more than 20 hours. It used 2.5 times as many tokens. This report comes from Anthropic's launch material. It shows the intended advantage, but it is not a promise for every project.
What the release actually shows
Anthropic's table reports 66.4% for Opus 5.5 on Terminal-Bench 4.0. Opus 5 scored 52.3%. GPT-6 Astra scored 57.9%. The settings were not identical. Opus 5.5 used xhigh effort, while Astra used high effort. Anthropic also says safety safeguards changed some results by routing tasks to other models. These figures are best read as reported comparisons, not a universal ranking.
Anthropic says Opus 5.5 set a new high on its automated behavioral audit. It says the model is less likely to take hard-to-reverse actions. It also reports stronger resistance to misleading instructions. Biology and cybersecurity work receive additional safeguards. Those protections matter, but they can affect benchmark scores.
What remains uncertain
Most evidence here comes from Anthropic's own tests and early testers. The company itself warns that benchmark margins can mislead. Real costs will vary with effort, caching, tools, and fallbacks. We still need broader independent testing. We also need more evidence from long, unattended jobs in real workplaces.
What to watch next
Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow soon. The important question is practical. Can teams run longer jobs reliably? Can they keep costs low while checking the results? Independent evaluations and real use will show whether Opus 5.5 changes how people use AI.
Claude Opus 5.5 aims to make long AI jobs cheaper
📰 Full story: Claude Opus 5.5 shifts the AI race toward cheaper, longer work
Anthropic says its new Claude model works faster and costs less than Opus 5.
Claude Opus 5.5
Anthropic's new AI model for difficult work.
token
A small piece of text used to count work and price.
benchmark
A test that compares models using the same tasks.
💡 The gist
- Anthropic, the company behind Claude, released a new AI model.
- It says Opus 5.5 is cheaper and faster than Opus 5.
- Hacker News noticed it, but attention is not proof.
Claude Opus 5.5 is the first model in the Claude 5.5 family. It targets coding, computer tasks, research, and business work. Anthropic says typical work costs about 40% less. It also says the model creates answers more than 30% faster.
AI services often count text with tokens. A token is a small piece of text. Opus 5.5 costs $4 per million input tokens. It costs $20 per million output tokens. Opus 5 cost $5 and $25. Many small savings can matter during long jobs.
One early tester used Opus 5.5 on a 200,000-line codebase. The task took less than three hours. Opus 5 took more than 20 hours. It used 2.5 times as many tokens. This is a company-reported test, not a promise for every project.
The release also reports safety improvements. Anthropic says Opus 5.5 did better on its behavior tests. It says the model is less likely to take hard-to-reverse actions. It also says the model resists some misleading instructions better. Safety rules can still change benchmark results.
The main Hacker News post got 1,147 points and 784 comments. Another post got 273 points and two comments. This shows interest from a technology community. It does not show that the model is correct.
The next question is practical. Can teams run long jobs reliably? Can they keep costs low? Claude Sonnet 5.5 and Claude Haiku 5.5 are expected soon. Independent tests and real work will show how much this launch changes AI use.
💬 Claude Opus 5.5: Hope and doubt from HN users
Anthropic says Opus 5.5 should sound clearer and work well for coding, computer tasks, and knowledge work. HN users report mixed experiences, and their reports are personal observations rather than controlled tests.
- Some users’ self-reports call Opus 5 too wordy or awkward. One early 5.5 test still found the same wordy style and said the main point was hard to find.
- The post says benchmark scores may not match daily work. A commenter suggests that messy old code, maintenance, and unclear requests are hard for simple tests to measure.
- Some users’ self-reports say the announcement page’s forced scrolling is annoying; Reduce Motion reportedly gives a simpler page.
- People also disagree about slowing frontier AI: some hear a safety message, while others hear vague wording that could help powerful companies shape regulation. These are opinions, not proof.
mature digest at 1034 comments (revision 3). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
Claude Opus 5.5 is a new computer helper
📰 Full story: Claude Opus 5.5 shifts the AI race toward cheaper, longer work
Anthropic says its new Claude can do long jobs with less money.
Claude Opus 5.5
Anthropic's new computer helper.
Hacker News
A website where people discuss technology.
Anthropic, the company that makes Claude, made Claude Opus 5.5. Claude is a computer helper that reads words and writes answers. It can also write instructions for computers. Anthropic says it uses about 40% less money than Opus 5. It also answers more than 30% faster. That can help when a job takes many steps.
Hacker News is a place where people discuss technology. One post got 1,147 points and 784 comments. Those numbers show attention, not truth.
Anthropic tested the model on coding and safety. Some tests had safety rules that changed the result. We still need more real-world checking. Sonnet 5.5 and Haiku 5.5 are coming soon.
💬 Is Opus 5.5 easier to talk to?
Anthropic says Claude Opus 5.5 should be clearer and better at coding and computer work. People who tried it did not all agree.
- Some users say the older model talked too much. One early test says 5.5 still hid the important part in many words.
- Good test scores may not show what happens in a big, messy real project.
- Some readers disliked the moving web page; a Reduce Motion setting reportedly made it easier. People also disagree whether slowing AI is mainly for safety or for business and regulation.
mature digest at 1034 comments (revision 3). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
💬 Claude Opus 5.5: Claimed polish, divided real-world reactions
Anthropic presents Claude Opus 5.5 as more natural and clearer, with strong results in agentic coding, computer use, and knowledge work. HN commenters disagree about everyday quality, the value of benchmarks, and the announcement page’s UX; reports of quality or bugs are individual users’ self-reports, not controlled evidence.
mature digest at 1034 comments (revision 3). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.