Loading today’s stories
Leadership1 distinct publisher3 min readUpdated
A new cost analysis puts OpenAI's frontier model at half Anthropic's price per benchmark task. The retry and cleanup arithmetic behind that number is less settled than the price sheet.
The Board Room · Leadership desk
Compiled by The Board RoomSomething wrong?How this is made
A cost analysis published by doit.com argues that the unit of AI procurement has moved from dollars per million tokens to expected cost per completed task, plus the cost of cleaning up wrong outputs that escape automated checks [1]. That matters because on the old unit the cheaper frontier vendor has flipped, and on the new one the ranking depends on numbers most buyers do not currently measure [2][6].
Start with the price sheet, since that is where most comparisons stop. Anthropic's Claude Opus 5 lists at $5 per million input tokens and $25 output; OpenAI's GPT-5.6 Sol lists at $5 and $30, which makes Anthropic the cheaper option on output list price [3]. The framing that Claude costs more per token is, at the frontier, obsolete [3].
It is also irrelevant, because tokens are not a common currency. Anthropic's tokenizer from Opus 4.7 onward produces roughly 30 percent more tokens for the same text, and Opus 4.8 and Opus 5 report about 1.88 tokens per English word against about 1.17 on GPT-5's o200k encoding [4]. Normalise the list prices to words and the advantage reverses: about $47 per million words of Claude output versus $35.10 for GPT-5.6 Sol, and $9.40 versus $5.85 on the input side [17][18]. Reasoning tokens are billed as output on both platforms, with OpenAI hiding them and Anthropic able to return them, so a 500-token visible answer can carry thousands of billed tokens [5].
Measured per task rather than per token, Artificial Analysis put GPT-5.6 Sol at $1.04 per Intelligence Index task in mid-2026, against $2.03 for Claude Opus 5 and $2.75 for Claude Fable 5 [2]. That is a 1.95x gap [14]. The reason the gap does not track list prices is that agentic work is a loop, not a call: Vantage models a representative 50-turn coding session at roughly 1 million input tokens and 40,000 output, a 25-to-1 ratio in which input is about 85 percent of the bill because the model re-reads an accumulating context every turn [7]. Gartner's March 2026 analysis found agentic workflows burn 5 to 30 times more tokens per task than a chatbot query [8]. A preprint by Bai et al. covering eight frontier models on SWE-bench Verified found agentic coding consumes up to 1,000 times more tokens than code chat, with up to 30x run-to-run variance [10]. Priced at list, that 50-turn session costs about $6.00 on Opus 5 and $6.20 on GPT-5.6 Sol [19].
The case for the pricier model is reliability. On tau-bench, Claude Opus 4.8 held far more of its single-attempt score across repeated runs and violated policy less than half as often as GPT-5.5 [6]. Since a failed trajectory still bills in full [13], and expected cost per completed task is attempt cost divided by success probability [16], a 1.95x cost premium needs roughly 1.95x the success rate to break even on retries alone [15]. The rest has to come from cleanup avoided, which is exactly the line item nobody instruments. Gartner's Will Sommer put the general risk this way: product chiefs "should not confuse the deflation of commodity tokens with the democratization of frontier reasoning" [9].
Watch whether vendors start publishing repeated-run reliability alongside price, and whether your own finance function can produce a retry count. Bai et al. note that the models themselves underestimate their token usage [11]; prompt cache reads cost 10 percent of input on both platforms, which changes long-session arithmetic more than any list-price move [12].
Follow any of these and your For You feed starts watching them — no settings page required.
build
Three frontier launches in a day, all pitched on price. Open weights set the ceiling.4 distinct publishers
build
OpenAI puts latency on the price list: 750 tokens/sec, gated by workload fit3 distinct publishers
invest
Speed becomes a SKU: OpenAI and Google put a separate price on latency3 distinct publishers
build
Grok 4.6 lands in Copilot two days after launch, and the model picker becomes a procurement problem1 distinct publisher
Ranked by verification strength, evidence, and original report placement.
A doit.com analysis argues cost per token is the wrong unit for agentic and production workloads, and that the right unit is expected cost per completed task plus the cost of cleaning up wrong outputs that escape automated checks.
Some fraction of tasks fail, and a failed trajectory still costs full price.
On raw per-task token cost in mid-2026, Artificial Analysis measured GPT-5.6 Sol at $1.04 per Intelligence Index task versus Claude Opus 5 at $2.03 and Claude Fable 5 at $2.75.
The analysis defines the cost of a single attempt as the sum over trajectory steps of billed tokens times their rates, and models attempts as succeeding independently with a given probability, so that expected cost per completed task is attempt cost divided by success probability.
Flagship list prices have converged: Claude Opus 5 costs $5 input / $25 output per million tokens and GPT-5.6 Sol costs $5 / $30, making Anthropic cheaper on output list price and rendering the older "Claude costs more per token" framing largely obsolete at the frontier.
On tau-bench, Claude Opus 4.8 held far more of its single-attempt score across repeated runs and violated policy less than half as often as GPT-5.5.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 18, 2026
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Named third-party numbers, one interested publisher, no corroboration
Every figure in the cluster comes from a single commercial blog post that attributes its numbers to identifiable outside parties — Artificial Analysis's Intelligence Index, Vantage's session model, Gartner's March 2026 analysis with a named analyst quote, and a citable arXiv preprint with released raw data. That attribution is better than anonymous assertion, but nothing here is independently confirmed inside the cluster, the pricing tables and the formal equations are missing from the supplied body, and the 'Cost per task, measured' section is truncated mid-sentence so the benchmark's measurement conditions cannot be inspected. The decisive quantities for the article's own thesis — per-attempt success rates and the cost of cleaning up leaked wrong outputs — are never measured.
Benchmarks and price sheets exist; practitioner uptake unevidenced
There is concrete, dated market plumbing behind the story: an index that publishes absolute dollar cost per task, tau-bench reliability runs, a SWE-bench Verified trajectory study with public raw data, and live pricing mechanics including Sonnet 5's introductory window, GPT-5.6 cache-write pricing and OpenAI's long-context surcharge. What is entirely absent is evidence that anyone has adopted the article's unit of account: no team, platform or vendor is shown reporting expected cost per completed task, no spend disclosures, and no deployment measuring retry or cleanup rates in production. Adoption is therefore scored on the surrounding measurement and pricing infrastructure, not on the methodology itself.
Price-sheet facts solid; the retry-and-cleanup case overreaches its data
The pricing and tokenizer findings are stated crisply and are internally consistent, so the gap is not in the sticker numbers. It is in the framing built on top of them. The article's strongest assertion — that reliability, retries and human cleanup are 'where most of the real money is' and that the cleanup term is 'frequently larger than the bill itself' — is supported by no measurement of leak rates, cleanup costs or per-attempt success rates for the models being compared; the reliability evidence offered is for Opus 4.8 versus GPT-5.5, one generation behind the priced pair. The flat-rate 50-turn session estimate also contradicts the article's own long-context surcharge for 1M-input requests on GPT-5.6, and holding token counts equal across vendors sits awkwardly against its own 1.88-versus-1.17 tokens-per-word finding. The overstatement is moderate and mostly analytic, not fabricated.
Cost-management vendor blog arguing for cost measurement
The single source is a company blog post about controlling LLM spend, published on a commercial site's /blog path with no methodology note or interest disclosure. Its central recommendation — stop comparing price sheets and start measuring expected cost per completed task, including invisible cleanup costs — is exactly the kind of conclusion that creates demand for spend measurement and optimisation services. Offsetting this, the piece is not promoting either model vendor: it hands the raw-cost win to OpenAI and the reliability win to Anthropic, cites outside benchmarks by name, and points readers to third-party raw data they can check. So the incentive is present and structural rather than hidden advocacy for a specific product being compared.
Directionally usable, single-sourced, partly self-contradictory
Confidence is limited by structure more than by content quality: one publisher, one document, several figures reported secondhand, a truncated body, and an internal contradiction between the flat-rate session estimate and the documented long-context surcharge. The pricing, tokenizer and caching mechanics are specific enough to act on with verification against vendor rate cards; the reliability-and-retry conclusion is not yet load-bearing because the success-rate and cleanup terms are unmeasured and the benchmark cited is a generation stale.