Journal du Net’s four-task comparison of OpenAI and Anthropic API costs found that the lowest bill did not always come with a successful run. In the article, published October 19, 2026, Claude Haiku 5.5 passed none of three Flask code-repair runs, while GPT-6 Luna passed one of three deduplication runs. The test date was not reported. Journal du Net’s account of the test
There is no overall provider winner in these results: the reported charges changed by task, and the pass counts matter. Journal du Net says it ran each model three times with high reasoning effort and prompt caching enabled. Its cost calculation for Anthropic included a cache write on the first pass and cache reads on the next two.
Four tasks, six models
The comparison covered a Flask bug fix, extracting data from 20 fictitious PDF invoices into JSON, answering 20 numerical questions about a 400-page SpaceX pre-IPO prospectus, and deduplicating 300 synthetic customer records that included 40 duplicates. For the Flask task, a fix counted as successful when it passed Flask developers’ added tests.
Reported charges and passes
The table shows Journal du Net’s reported dollar charges for each task and the number of successful runs out of three. These are task results from its test, not a general price card.
| Task | GPT-6 Astra | Claude Opus 5.5 | GPT-6.1 Sol | Claude Sonnet 5.5 | GPT-6 Luna | Claude Haiku 5.5 |
| Flask bug repair | $0.4278; 3/3 | $0.2534; 3/3 | $0.0730; 3/3 | $0.1217; 2/3 | $0.0048; 3/3 | $0.0398; 0/3 |
| Invoice extraction | $0.4023; 3/3 | $0.2405; 3/3 | $0.0791; 3/3 | $0.1284; 3/3 | $0.0041; 3/3 | $0.0088; 3/3 |
| SpaceX prospectus questions | $2.9275; 3/3 | $1.0770; 3/3 | $0.5369; 3/3 | $0.5358; 3/3 | $0.0294; 3/3 | $0.1458; 3/3 |
| Customer-record deduplication | $0.6184; 3/3 | $0.3451; 3/3 | $0.1126; 3/3 | $0.1673; 3/3 | $0.0061; 1/3 | $0.0105; 3/3 |
GPT-6 Luna had the lowest reported charge in all four tasks, but its deduplication result was one successful run out of three. Haiku’s code-repair charge was $0.0398, with no successful runs. A small per-task bill and reliable completion are separate measures in these results.
Token counts and prompt length change the bill
For the same test inputs, Journal du Net reported different input-token counts for the two providers. The OpenAI and Anthropic counts were 77,979 and 120,242 for code repair; 5,399 and 7,965 for invoices; 362,517 and 592,569 for the prospectus; and 11,502 and 16,745 for deduplication. Those counts describe these tasks, not a universal conversion between text and tokens.
Anthropic’s pricing page sets a 100,000-token prompt threshold for Claude Haiku 5.5. At or below that threshold, input costs $0.10, cache reads $0.01, cache writes $0.125, and output $0.50 per million tokens. Above it, the listed rates are $0.50, $0.05, $0.625, and $2.50, respectively, a fivefold increase in each category. Anthropic’s API pricing also lists Claude Opus 5.5 at $4 per million input tokens, $0.20 for cache reads, and $20 for output; Claude Sonnet 5.5 is listed at $2, $0.10, and $10, respectively.
Model names on OpenAI’s pricing page
OpenAI’s API pricing page lists GPT-6 Astra and GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna. Journal du Net’s comparison instead names GPT-6.1 Sol and GPT-6 Luna alongside GPT-6 Astra. The page lists GPT-6 Astra short-context rates of $10 per million input tokens, $1 for cached input, and $50 for output. OpenAI’s API pricing
For a new workload estimate, use the rate for the exact API model and token category involved. Journal du Net’s task charges show what it reported for these four inputs and runs; they do not substitute for pricing a different workload against its own token use and completion needs.
