Haiku 5.5 Lists at $0.10. OSWorld Hit 72.4%.

The list price people will paste into a spreadsheet is the short-prompt price. Anthropic’s Haiku 5.5 page puts input at $0.10 per million tokens and output at $0.50 per million, for prompts up to 100,000 tokens. Over that line the same table goes to $0.50 in and $2.50 out. SiliconANGLE’s October 7 story leads with a different fraction: roughly a quarter of what Haiku 4.5 costs to run. Both numbers are in the source set. They are not the same cut. If your note says “90 percent off” and your other note says “75 percent off” and your headline says “one quarter,” you have three sentences that can all be traced, and only one of them is the sticker.

We already wrote a price cut that was not a capability jump in Opus 5.5, Sol, and the halved rates. We wrote a benchmark a frontier model barely cleared in the 34.8 percent on Argo-Bench. We wrote a score that was a rank, not a nationality trophy, in Mistral Large 4 at 38. Haiku 5.5 is a fourth shape: a small-model sticker, an effort dial that moves both the score and the bill, and an average saving that includes a tokenizer change. The model id on the Claude Platform is claude-haiku-5-5.

The sticker, the average, and the quarter

Haiku 4.5’s list price, as SiliconANGLE states it, is $1 per million input tokens and $5 per million output. Anthropic’s comparison table matches that column: input $1.00, output $5.00. Against those stickers, $0.10 and $0.50 are a 90 percent cut on both sides, for prompts at or under 100,000 tokens. Anthropic says about 90 percent of requests to the older Haiku fell in that bucket. MarkTechPost’s October 7 piece uses the same 90 percent line and the same $0.10 / $0.50 pair.

Longer prompts get a 50 percent discount, not 90. SiliconANGLE says that in the body, after the “quarter” framing. Anthropic’s footnote is the one to pin on the wall: priced 90 percent lower than Haiku 4.5 for requests up to 100,000 tokens, and 50 percent lower for requests over 100,000 tokens. The table’s long-prompt cells are $0.50 input and $2.50 output. Half of $1.00 is $0.50. Half of $5.00 is $2.50. The long-prompt sticker is the 50 percent sentence. It is not a rounding of the short-prompt sticker.

The 75 percent figure is an average running-cost claim, not a cell in that table. Anthropic says Haiku 5.5 costs around 75 percent less to run than Haiku 4.5, on average. The footnote says the calculation also accounts for a new tokenizer, similar to Sonnet 5.5’s and Opus 5.5’s, which uses slightly more tokens per task. SiliconANGLE repeats that. A 90 percent list cut, applied to a mix of short and long prompts, then adjusted because each task now spends a few more tokens, can land near 75 percent. “Around” is their word.

“Roughly a quarter of the running cost” is the same neighborhood as “75 percent less,” said from the other side. Quarter-of-old versus 90-percent-off-the-sticker will diverge the moment your prompts sit over 100,000 tokens, or the moment the tokenizer inflation is larger on your workload than on theirs. Quote the cell you will actually be billed on. Quote the 75 percent only if you are describing Anthropic’s average, and say it is theirs.

Cache prices sit on the same table, and they are easy to skip because the headline is about input and output. For Haiku 5.5, cache reads are $0.01 per million under 100,000 tokens and $0.05 over that line. Cache writes are $0.125 and $0.625 on the same split. Haiku 4.5’s column shows $0.10 cache reads and $1.25 cache writes. Sonnet 5.5’s column shows $0.10 cache reads and $2.50 cache writes, which is the post-cut Sonnet read price, not the old $0.20. If your agent loop is mostly cached prefix, the $0.10 input sticker is not the line that dominates the invoice. The $0.01 cache read is.

MarkTechPost adds two capacity numbers the Anthropic page did not put in the paragraphs we transcribed: a 1 million token context window, and up to 128,000 output tokens. I am attributing those to MarkTechPost, not pretending they were in the hero sentence on anthropic.com. The same piece says the effort setting defaults to medium. Anthropic’s page says Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, and the charts show Low through Max. I did not find the word “default” in the Anthropic text we pulled. If you set nothing and assume medium, you are following MarkTechPost, not a sentence I can point to on the product page.

Knowledge cutoff, again from MarkTechPost’s comparison table: June 2026 for Haiku 5.5, May 18, 2026 for GPT-6 Luna. A cutoff is not a training-data affidavit. It is the date they printed. Do not promote it into “the model knows events through June” without their eval note.

Where Haiku 5.5 sits against Luna and Sonnet

The comparison that will get clipped is Haiku versus GPT-6 Luna, because the short-prompt sticker matches. SiliconANGLE says the $0.10 and $0.50 rates are what OpenAI charges for Luna. MarkTechPost says the same, and adds that Luna is cheaper above 100,000 tokens. Same sticker under 100,000 is not the same sticker over it. If your job is long-context extraction, the “same price as Luna” line is the wrong row.

Anthropic’s published table, which SiliconANGLE also walks, has Haiku 5.5 ahead of Luna on the tests where both have a score. Here are the cells on Anthropic’s page. Haiku 5.5, then Haiku 4.5, then Luna, then Sonnet 5.5.

GDPval-AA v2.1: 1620, 735, 1437, 1840. AA-Briefcase v1.1: 1578, 614, 1336, 1824. OSWorld 2.1, offline subset: 72.4 percent, 15.7 percent, 48.9 percent, 83.9 percent. Humanity’s Last Exam, no tools: 45.9 percent, 10.2 percent, no Luna cell, 56.9 percent. HLE with tools: 57.4 percent, 18.7 percent, no Luna cell, 64.5 percent. Terminal-Bench 4.0: 39.2 percent, 0.0 percent, 16.4 percent, 70.6 percent. FrontierCode 1.1 Main: 46.4 percent, no Haiku 4.5 cell, 42.4 percent, 52.1 percent at Xhigh for Sonnet. Chartography, no tools: 46.4 percent, 6.4 percent, 29.1 percent, 61.6 percent.

Luna is missing on both HLE rows. “Ahead on every shared test” can be true and still hide two empty cells. SiliconANGLE’s “six tests” is their count of the overlapping ones. Use the table if you are going to argue, not the count.

OSWorld is the number in the headline because 72.4 against 15.7 for Haiku 4.5 is a large jump on a computer-use task, and 72.4 against Luna’s 48.9 is the peer comparison. Sonnet 5.5 is at 83.9 on the same offline subset. Haiku did not pass Sonnet there. It passed the previous Haiku and it passed Luna. Those are different sentences. A procurement note that says “Haiku 5.5 leads computer use” has dropped the Sonnet column.

Terminal-Bench 4.0 is the column that should stop a coding-lead migration. Haiku 5.5 at 39.2 percent, Luna at 16.4, Haiku 4.5 at 0.0, Sonnet 5.5 at 70.6. Anthropic’s own caption says Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding of the kind Terminal-Bench measures. Haiku is aimed at narrower tasks that might otherwise have gone to a larger model: summaries, compaction, database queries, classification, and subagent work under Opus or Sonnet. SiliconANGLE dates the family: Opus 5.5 on September 22, Sonnet 5.5 on September 28, Haiku 5.5 on October 7. Three models, three jobs. The October launch does not retire the September ones.

FrontierCode is the place a vendor quote will try to overrule that. Cognition says that with Haiku 5.5 as the sidekick in Devin Fusion, and Opus 5.5 as the lead, Fusion holds a FrontierCode score of 66.2. That 66.2 is a two-model system. It is not Haiku’s 46.4 on FrontierCode 1.1 Main. If a slide shows 66.2 under the word Haiku, it has moved the lead model’s work onto the cheap model’s name.

Effort is a price, not a personality

The charts on Anthropic’s page plot accuracy against cost at five effort settings. The aria labels are specific enough to quote, and the max points match the table, which is how I am assigning each series.

On OSWorld 2.1 offline, Haiku 5.5 goes Low 42.0 percent at $0.07, Med 53.3 percent at $0.13, High 61.3 percent at $0.18, Xhigh 67.6 percent at $0.28, Max 72.4 percent at $0.61. The 72.4 in the comparison table is the Max point, not a free property of the model id. If your integration leaves effort at medium, MarkTechPost’s default, you are buying something nearer 53 percent on that chart, at $0.13, not the headline 72.4 at $0.61. I cannot see your request logs. I can see that the page printed both dots.

On GDPval-AA the same five steps are 1125 at $0.01, 1277 at $0.03, 1420 at $0.09, 1513 at $0.27, and 1620 at $0.87. The table’s 1620 is Max. A medium run at 1277 is still far above Haiku 4.5’s 735, and it is not the number you should put next to Sonnet’s 1840.

On Humanity’s Last Exam without tools: 30.9 percent at $0.003, 35.5 percent at $0.007, 39.7 percent at $0.02, 44.1 percent at $0.06, 45.9 percent at $0.21. Again the table’s 45.9 is Max. The cheap end of this chart is three-tenths of a cent per whatever unit the chart is using for that dot. I am not converting the chart’s cost axis into a monthly invoice. The shape is the point. Effort is how they want you to buy less intelligence on purpose.

Haiku 4.5 on the OSWorld chart’s max label sits at 15.7 percent, matching the table. On a Terminal-Bench cost chart, one Haiku 4.5 max label reads 0.0 percent at $0.79. That matches the 0.0 cell. Spending more effort on the old small model did not, in the label they printed, produce a Terminal-Bench score. That is a reason the new model’s 39.2 matters, and a reason not to describe 39.2 as “coding solved.”

Vendor quotes are not a second scoreboard

Anthropic prints customer lines. SiliconANGLE prints one of them with numbers the product page’s excerpt did not carry. Treat them as vendor-supplied, not as a lab you can re-run from this article.

Asana’s Aaron Vinh, staff software engineer, ran Haiku 5.5 through the eval suite for AI Teammates: triaging bugs, setting up projects, searching portfolios for high-risk or overdue work. SiliconANGLE adds the figures: task-completion latency more than 30 percent lower than the model Asana uses today, inference on each agent turn up to 2.5 times faster, and Vinh’s “noticeably snappier.” “The model we use today” is not named in the sentence. A 30 percent latency cut against an unnamed incumbent is a customer claim with a missing baseline. Useful as a direction. Not useful as a row next to OSWorld.

HubSpot’s Ze’ev Klapow says simulated portals, CRM tasks such as reporting on deals, and a best-yet score of 92.8 percent averaged over three runs. One audit task is mentioned and then the excerpt breaks off. 92.8 on their suite is not 92.8 on OSWorld. Do not import it.

AlphaSense’s Daniel Campos says Ask in Document does about 8 million calls a week. On 400 queries, Haiku 5.5 scored 0.84 against Haiku 4.5’s 0.76, and he calls the difference statistically significant. Four hundred queries is a real n for a claim of that size. It is their query set. The 8 million is volume, not a quality sample.

Box’s Yashodha Bhavnani says Haiku 5.5 scored 11 points higher than Haiku 4.5 at about half the latency, in early testing on enterprise content. Eleven points on an unnamed internal score is eleven points. Half the latency is a speed claim. Neither one is the GDPval 1620.

Rogo’s line, on the Anthropic page, is the job description rather than a score: short, high-volume work, lookups, subagents, summaries, including a subagent that pulls a segment-revenue line out of a 10-K while a larger model builds the deck. That is the positioning in a customer’s mouth. It matches Anthropic’s own “not the coding lead” caption. I would rather quote Rogo for the job split than quote Cognition’s 66.2 without the lead model attached.

The Sonnet cache cut is a separate invoice

The October 7 launch also halves cache reads on Sonnet 5.5, from $0.20 per million tokens to $0.10. Anthropic says cache reads are a large share of token consumption, so the cut reduces the cost of Sonnet 5.5 on most agentic tasks by around 20 percent. “Most” and “around” are doing work. If your Sonnet traffic is uncached, you do not get that 20 percent. If it is almost all cache reads, you get more than 20 percent off the cache line and the blended number depends on your mix.

Their Terminal-Bench 4.0 cost chart prints the same accuracies at two cache prices. Sonnet 5.5 at the new $0.10 cache read: Low 20.0 percent at $0.62, Med 28.8 percent at $0.68, High 43.0 percent at $1.46, Xhigh 61.5 percent at $4.34, Max 70.6 percent at $10.44. The old $0.20 cache series keeps the same percents and raises the Max dot to $14.15. The accuracy did not move. The Max bill on that chart dropped from $14.15 to $10.44. That is the picture they want next to the 20 percent sentence. It is a chart unit, not your monthly Sonnet invoice. It is still the cleanest demonstration in the pack that this particular cut is price, not score.

A second commercial change is not a model at all. Anthropic says that this week it will roll out a monthly API credit on the Claude Platform for Max and Team subscribers. Max 5x gets $100 a month. Max 20x gets $200. Team subscribers get up to $500, pooled across users. Those credits are for the platform, not a discount on the Haiku sticker itself. A team that burns the pooled $500 on Max-effort Haiku calls will learn the effort chart faster than a team that reads it. $0.61 on the OSWorld Max dot is not $0.61 per million tokens. I am not collapsing those units. I am saying a credit pool without an effort cap is a pool.

Python and TypeScript SDKs are getting computer-use and browser-use support in beta. Anthropic points at Haiku 5.5 for those tasks because of speed, capability, and price, and points at the platform docs. Beta support is not the same thing as the 72.4 percent offline OSWorld number. One is an API surface. One is an eval. Shipping the SDK beta does not move the eval.

Safeguards, and where to call it

Alignment, in Anthropic’s wording: major improvements on almost all of their alignment evals relative to Haiku 4.5, far fewer instances of misaligned behavior, lower willingness to cooperate with misuse. The system card is the document they say holds the method. SiliconANGLE’s paraphrase matches the “far fewer” line. I have not re-scored the system card here. “Far fewer” without a table in the blog post is a direction, not a rate.

Cybersecurity safeguards are more restrictive than Haiku 4.5’s and somewhat less restrictive than the ones on other recent models. They allow a wider range of defensive work than Sonnet 5.5’s safeguards, and they still block penetration testing and other techniques more likely to be used by attackers. Biology safeguards match Sonnet 5, Sonnet 5.5, and Opus 5: research questions allowed, requests judged likely to cause harm restricted. Wider biology or cyber work goes through the Life Sciences Verification Program or the Cyber Verification Program. If your use case is offensive security, this paragraph is the stop sign, not a puzzle to route around with a prompt.

Availability, from Anthropic: all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure, and claude-haiku-5-5 on the Claude Platform. MarkTechPost’s list adds Amazon Bedrock, Microsoft Foundry, and Claude Platform on AWS, and says generally available as a hosted API. The operational fact is that this is a hosted model id, not a weight file. Unlike the Mistral preview we wrote about on October 8, there is no October 27 weight date in these posts, because nobody is promising you the weights.

Speed has a footnote that headlines will drop. Haiku 5.5 is their fastest model to date at each model’s standard speed. It runs less quickly than Opus models in Fast Mode. “Fastest Haiku” and “faster than Opus Fast” are not the same claim. The footnote exists so you do not make the second one.

If you are picking a default for high-volume classification this week, the short-prompt cell is $0.10 in and $0.50 out, the context figure to cite is MarkTechPost’s 1 million, and the OSWorld number to cite is 72.4 percent only if you are paying for the effort setting that produced it. Sonnet remains the Terminal-Bench column at 70.6 percent. The 75 percent average is Anthropic’s blend, tokenizer included. Write the cell. Leave the quarter for the headline writers who have not opened the table.