Anthropic's cheapest new model just beat OpenAI's cheapest new model by 23.5 25 points on a test of operating a computer. The two cost the same. Anthropic reports Claude Haiku 5.5 scored 72.4% 01 on the OSWorld 2.1 offline subset. GPT-6 Luna scored 48.9% 02.
Anthropic's release page lists Haiku 5.5 at $0.10 11 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens, and $0.50 and $2.50 above that. moelueker.com's launch review notes that, up to 100K tokens, that is exactly GPT-6 Luna's sticker price. Same price, very different results on agent work.
The cheap tier is now the front line. OpenAI is rolling Luna out to ChatGPT Free and Go users from October 9, while paid tiers get GPT-6 Sol. For millions of people, the budget model is the only one they will ever touch.
One headline from this week needs correcting first. We could not find a Microsoft announcement or Azure pricing page confirming that, so treat the Microsoft 75% 09 figure as unverified. The 75% cut we can source belongs to Anthropic, and it tells a more useful story.
Same sticker price, different finished jobs
The sticker gap is the distance between what a model costs per token and what it costs per job that actually finishes. This week's numbers show it twice.
What a finished agent task actually costs on the cheap tier
The first case is Haiku 5.5 against Luna. On Anthropic's OSWorld figures, Haiku needs about 1.38 26 attempts per success (1 divided by 0.724). Luna needs about 2.04 27 attempts (1 divided by 0.489). Same token price, and Luna burns roughly 48% 28 more attempts for every finished task.
The second case is Anthropic's own price cut. ppc.land reported that the rate card fell from $1.00 13 and $5.00 on Haiku 4.5 to $0.10 and $0.50 on Haiku 5.5 for prompts up to 100,000 tokens, with larger prompts costing $0.50 and $2.50. Anthropic's release page says the model costs "around 75% 09 less to run" on average.
So where did the missing 15 points go? Projedefteri.com has the answer. Haiku 5.5 ships with a new tokenizer, the piece that chops text into billable units. It also has a long-prompt tier: any request over 100,000 tokens 14 costs five times more, at $0.50 in and $2.50 out.
Every agent bill comes down to three things. First, the price per attempt: tokens in and out, at whichever tier the prompt lands in. Second, the success rate, meaning the share of attempts that finish the job. Divide the attempt price by it. Third, cleanup: the retries and human fixes on the runs that failed.
Brand names show up in none of them.
Reading Anthropic's table like an ops lead
Picture paying a courier per trip. The logo on the van doesn't matter. You care about parcels delivered per trip, and how many came back damaged.
Anthropic's comparison table gives you the delivery rates, and the margins swing hard by job type. Haiku 5.5 leads Luna 72.4% 25 to 48.9% on OSWorld 2.1. On FrontierCode 1.1 29 the lead shrinks to 4 points, 46.4% against 42.4%.
That spread is the whole lesson. A 23.5-point 25 lead on clicking through desktop apps says little about a coding task where the gap is 4. Pick the benchmark that looks most like your job and ignore the rest.
The bigger sibling still matters. Aiidelist.com, reporting Anthropic's numbers, puts Claude Sonnet 5.5 at 83.9% 04 on OSWorld's offline subset (Anthropic's headline score was 80.1%) and 70.6% on Terminal-Bench. Haiku's 39.2% 30 on the terminal test is a bit over half of Sonnet's. Projedefteri.com's advice follows from that: use Haiku for summaries and browser work, and keep it out of the lead seat on complex agentic coding.
Anthropic's own launch page says the same thing more gently. It pitches Haiku 5.5 as a subagent, a model handed one slice of a bigger job, working under Opus 5.5 or Sonnet 5.5. Anthropic also halved Sonnet 5.5's cache-read price, which it says makes Sonnet around 20% 21 cheaper on most agentic work. So the escalation model got cheaper on the same day the cheap model got better.
Take a job that sends 1 million 31 input tokens and gets back 200,000 output tokens. Under 100,000 tokens per request, that's $0.10 plus $0.10, so $0.20.
Push the same volume through prompts over 100,000 tokens 32 and it costs $0.50 plus $0.50, so $1.00. How fast your agent's context grows decides which tier you live in.
Retries widen the gap. Assume both models spend $0.20 33 per attempt. Haiku costs about $0.28 per finished task ($0.20 times 1.38). Luna costs about $0.41 34 ($0.20 times 2.04). The equal-tokens-per-attempt assumption is ours, and real runs will vary.
Every Haiku-versus-Luna number above comes from Anthropic's own evaluation setup, and that is the weak point. BenchLM.ai's comparison page, updated October 7, says the public evidence has no benchmark result shared by both models. In BenchLM's words, that evidence "does not support a quality verdict."
Nobody outside Anthropic has shown whether the 23.5-point 25 gap survives a different harness with Luna run at its best settings. We think the table justifies a test. It doesn't justify a migration. The test is cheap, and the next section shows why it keeps paying.
Hidden costs behind identical sticker prices
Anthropic's OSWorld lead lacks outside confirmation.
Every Haiku-versus-Luna number comes from Anthropic's own evaluation setup. BenchLM.ai found no benchmark result shared by both models, so the gap justifies a test and still needs an independent harness.
Growing context quietly multiplies the bill.
Haiku 5.5 requests over 100,000 tokens 14 cost $0.50 in and $2.50 out. The same 1 million 32 input and 200,000 output tokens cost $0.20 31 below the line and $1.00 above it.
Retries and review become most of the bill.
If small-model running costs keep falling about 75% 36 per generation, a $1.00 job drops to about $0.016 by 2031. Failed runs and human fixes do not shrink at that pace, and benchmarks like IFBench are already being retired.
2031. Count the release dates. Projedefteri.com lists Opus 5.5 on September 22, Sonnet 5.5 on September 28 and Haiku 5.5 on October 7. Anthropic refreshed its entire lineup in 15 days 35.
Measurement moved just as fast. On June 15, 2026 24, Artificial Analysis announced version 4.1 of its Intelligence Index. It reweighted the index toward agentic tasks and added three new metrics, led by Cost per Task. The other two are Time per Task and Tokens per Task. Four months later, moelueker.com's Haiku launch cover charts OSWorld score against cost per task. The scoreboard is moving toward the finished job.
The napkin math goes like this. Anthropic says one small-model generation cut average running cost by about 75% 36. Suppose that happens three more times by 2031. A $1.00 job becomes $0.25, then about $0.06, then about $0.016.
Straight lines break, so hold that loosely. The direction is still right. As token spend falls toward pennies, failure costs don't fall with it. Retries and human review end up as most of the bill.
Benchmarks age too. In the same June update, Artificial Analysis dropped IFBench from its index because frontier models had saturated it. Sonnet 5.5 already sits at 83.9% 04 on the OSWorld subset. By 2031, the test you trust this week may be retired.
Put the two trends together and the bet is lopsided. A private test set of your own tasks costs a few dollars of tokens and a day of setup. Choosing by brand costs you on every run, every month, across every launch you skip evaluating. And a test set you keep rerunning compounds: each new model takes an hour to judge, while competitors are still reading launch posts.
Score twenty real tasks before Friday
The goal is one spreadsheet that tells you which model to send each job to, with the cost of a finished task beside each name.
Step one: define done. Pick one job your agent runs daily, such as filing a support ticket or filling a web form. Write one sentence that says what a finished run looks like. Pull 20 real examples, and make sure at least 5 are ones that went wrong before.
Step two: run the bake-off. Send all 20 through Haiku 5.5 (API name claude-haiku-5-5), GPT-6 Luna and whatever model you use today. Keep the prompt, tools and stop rule identical across all three. Log input tokens, output tokens, pass or fail, retries and minutes a person spent fixing the output.
Step three: price each attempt. Divide input tokens by 1,000,000 and multiply by the input price. Do the same for output. Add the two to get cost per attempt.
Step four: price each finished task. Divide cost per attempt by the pass rate. Pass rate means the share of runs that met your one-sentence definition of done. Then add the human minutes, multiplied by what an hour of that person costs you.
Step five: flag the expensive tier. Mark every Haiku request that crossed 100,000 tokens 14. Those cost five times more. If your agent keeps crossing the line, trim its context before you blame the model.
Step six: write the routing rule. Send the job to the cheapest model whose pass rate clears your bar. Anything that fails the check escalates to Sonnet 5.5. Score the cost of that escalation path as well.
Expect things to break. Browser sessions will log out, tool calls will time out and some runs will hang. Count every one of them as a fail, because production will. Then put a reminder on the sheet: rerun all 20 tasks the day any vendor changes a price or ships a model. On this year's evidence, that could mean every few weeks.
Score twenty real tasks and write a routing rule
- Define done and pull 20 examples. Pick one daily agent job, write a one-sentence finish line, and include at least 5 runs that failed before.
- Run an identical bake-off. Send all 20 through Haiku 5.5, GPT-6 Luna and your current model with the same prompt, tools and stop rule, logging tokens, pass or fail, retries and human fix minutes.
- Price the finished task. Divide cost per attempt by pass rate, add human minutes at their hourly cost, and flag every Haiku request over 100,000 31 tokens before routing jobs to the cheapest model that clears your bar.
Price the finished job, then pick the model.
Haiku 5.5 and GPT-6 Luna share a sticker price and land far apart on Anthropic's agent tests. Those results are self-reported, so we treat them as a reason to run a test of our own. Twenty real tasks, a spreadsheet and a few dollars of tokens tell you which model finishes your work for less. Rerun it every time a vendor ships a model or changes a price, which this year means every few weeks.