ClarityBeta

Leadership & Executive

The Board Room

8 min read · 1,700 words · 3 moves to watch
FRIDAY, OCTOBER 9, 2026
The Signal

Haiku 5.5 ties GPT-6 Luna on the price sheet and loses on the bill.

The five-point lead at maximum effort is real, and it costs roughly three times the rival's tokens to earn. The tokenizer then counts the same text 25-30% heavier, and prompts past 100,000 tokens pay five times the base rate. That means any cost model you ran on list price will hand the long-context jobs to the model that bills them worst.

In Play

  1. Small-Model Price War Hides in the Fine Print

    Anthropic priced Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens, the same as OpenAI's GPT-6 Luna. Artificial Analysis scores Haiku higher, 43 to 38. Prompts over 100,000 tokens pay five times the base rate, and Haiku's tokenizer counts the same text as roughly 25-30% more tokens. If your cost model uses list price, it will pick the wrong vendor for long-context work.

  2. Frontier Lab Revenue Was Overstated

    The Financial Times puts OpenAI's annualized revenue near $50B at the end of September, not the widely reported $70B. Nvidia, CoreWeave and Nebius all fell on the report. Many vendor-viability models and capacity plans are anchored to lab revenue. That revenue is unaudited, estimated as one month times 12, and calculated differently by OpenAI and Anthropic.

  3. OpenAI's Outside Oversight Is Thinning

    OpenAI fired three safety researchers, including Tomek Korbak. Korbak was its main technical contact with the outside auditor METR while METR reviewed July's agent containment incident. The researchers say they were fired for putting safety first. OpenAI reportedly says they mishandled confidential information. Your vendor reviews probably score model safety, which still looks strong, but not whether outside auditors keep their access.

  4. Agents Become a Customer Class

    Agents generated 48% of usage of Cloudflare's Wrangler command-line tool in one week. Meta and Sierra launched the Personal Agent Protocol with Shopify, Stripe and Walmart as partners. Stripe and Cloudflare are each building ways to charge agents per request. Your seat-based pricing and employee-based identity model were designed for human buyers.

  5. AI Returns Accrue to Builders

    A third of The Information's tech-forward subscribers say AI returns a multiple of what they spend. A year ago, MIT found that only 5% of custom enterprise tools reached production. Separate research finds just 7% of companies have scaled AI, so the survey probably reflects organizations that build software. Respondents also report using Microsoft Copilot less even as seats grow, which suggests unused seats in your licensed tools.

Deep Dives

Haiku 5.5's Price Sheet Is Not Its Bill

Matching a rival's per-token price means little when a model reasons longer, counts text differently and charges five times as much past a context line.

Where the discount leaks out

Three mechanics sit between the headline price and your invoice. The first is token burn. Artificial Analysis found that Haiku 5.5 needs about 55,000 output tokens to match GPT-6 Luna's index score of 38. Luna needs 50,000. Haiku's five-point lead at maximum effort costs about 162,000 tokens, roughly three times what Luna uses at the same per-token price.

The second is counting. Anthropic's new tokenizer counts the same text as roughly 25-30% more tokens. Reasoning defaults to medium and cannot be fully turned off. The third is the context cliff. Above 100,000 tokens, Haiku charges $0.50 for input and $2.50 for output. Haiku matches Luna's list price only below 100,000 tokens, so for longer prompts Luna can be cheaper at list price.

Independent evaluators already see the effect. Vals found that Haiku 5.5 costs more per test than Haiku 4.5 on every benchmark the two share. On one legal research task it took 59 reasoning steps where its predecessor took 17. This is how Cursor can call the model "10x cheaper" for short coding requests while Anthropic claims "75% cheaper on average." Both can be true. The shape of your workload decides which one you get.

Counting requests instead of dollars

Anthropic says about 90% of Haiku 4.5 requests fit under the 100,000-token line. That figure counts requests, not dollars, and long prompts carry most of the tokens. Take an illustrative mix of 90 short requests at 10,000 tokens and 10 long ones at 300,000. The long requests make up about 94% of input spend. If your agents load whole documents into context, the headline cut may barely touch your bill.

Why it is still worth taking

None of this makes Haiku 5.5 a bad buy. It leads its tier on agentic coding, scoring 33% on Terminal-Bench against Luna's 13%. It also beat Sonnet 5 on Devin's FrontierCode at about one-eighth the cost per task. Anthropic halved Sonnet 5.5 cache reads to $0.10 per million tokens, which it says makes long-running agent work about 20% cheaper. GitHub Copilot and Devin already run Haiku as the worker beneath an Opus or Sonnet lead.

One gap is temporary but real. An over-refusal bug currently holds Haiku's automation score at 35%, against 53-60% for its rivals, until Artificial Analysis re-runs the test.

Expect the price war to continue. Observers are openly calling on OpenAI to respond. The Information also reports, in a headline without detail, that Microsoft cut its internal Claude spending by about a third, so even the largest buyers shop around. The advantage goes to whoever can re-benchmark fastest. A price cut only helps companies that can evaluate and re-route a workload within weeks.

A model that matches a rival's price per token can still cost three times as much per finished task.

The smart move

Make model selection a recurring, measured decision owned by one team rather than a one-time purchase. Price every workload on cost per completed task at its real context length. Set context and effort limits by policy instead of leaving them to individual engineers. Keep contracts short enough to capture the next price cut.

What to do

  1. Commission a 30-day bake-off of Haiku 5.5, GPT-6 Luna and Gemini on your top three high-volume AI workloads, scored on cost per completed task at real context lengths, ending in one decision memo.

  2. Issue a company-wide AI usage policy this quarter: keep worker calls under 100K tokens, set a default effort level for each workload class, and allow maximum effort only by exception.

  3. Cap new model commitments at 12 months with price-protection clauses, and keep two vendors live in production for every critical workload.

Two Corrections to Your OpenAI Vendor File

These developments weakened both the revenue figure behind vendor-viability models and the outside audit behind safety assurances, and most scorecards track neither.

A run rate built from leaks

The revenue gap comes from how the figure was calculated. The Information traces the overstatement to "clunky calculations" by investors and news outlets. OpenAI has restated nothing. On the usual convention of one month's revenue times 12, the difference is roughly $4.2B versus $5.8B a month. Separate reporting says part of the gap came from investors adjusting OpenAI's figures to match Anthropic's practice of counting sales made through cloud partners. The two labs define revenue differently, so a lab-to-lab comparison in vendor-viability work is probably comparing unlike figures.

The market reaction matters more than the number. A measurement correction at one private company moved the leading chipmaker and two GPU cloud providers on the same day. Compute suppliers' cost of capital now rides partly on OpenAI's story, whether or not a buyer uses OpenAI. Fifty billion dollars is still an extraordinary run rate, and demand has not collapsed; what has gone is outsiders' ability to verify it. Until the labs file audited financials, revisions will keep arriving the way private-market numbers always have, through investor decks and leaks.

An audit relationship under strain

The firing dispute is unresolved. The researchers say they were dismissed for "prioritizing safety over the near-term interests of OpenAI as a corporation." OpenAI reportedly says they mishandled confidential information. Neither account has been verified. The consequences do not wait for a verdict. Korbak fears OpenAI will use the firings to step back from METR, and Neel Nanda warns that firing staff over good-faith judgment will chill outside safety work.

The researchers also warn that the industry could lose oversight of AI reasoning, which is the ability to read a model's step-by-step working before it acts. Apollo argues that testing only the finished model could not have caught July's behavior, because the behavior emerged earlier in development. The same week, a senior OpenAI safety lead resigned publicly and called the culture "broken," and Common Sense Media labeled ChatGPT for Teens an "unacceptable risk." A skeptic would call this one company's bad month. Anthropic's public safety resignation a month earlier says otherwise.

Why vendor scorecards miss both

Product-level safety scores point the other way. GPT-6.1-Sol tops Arena's new Alignment Index at 87.9, ahead of Claude Opus 5.5 at 83.2. Model behavior and institutional governance are moving in opposite directions, and most vendor reviews measure only the first. Revenue gets the same treatment. Procurement teams cite run rates as proof a vendor will last and rarely ask how those figures were calculated.

GPT-6.1-Sol leads the Alignment Index at 87.9 while OpenAI's relationship with METR, its outside auditor, is fraying.

The smart move

The defensible position is to score frontier vendors on two separate axes, capability and governance, and to put the governance terms in the contract. The terms that matter most are access for third-party evaluators and timelines for disclosing incidents. A direct answer on whether new architectures make reasoning harder to monitor belongs in writing as well. The cost is slower procurement this quarter. The payoff is fewer surprises at renewal, and in the meantime lab revenue headlines deserve little weight in internal plans.

What to do

  1. Add governance criteria to every frontier-vendor scorecard and renewal this quarter: third-party evaluator access, incident-disclosure timelines, reasoning-monitorability commitments and how the vendor defines revenue. Score these separately from capability.

  2. Re-run every compute reservation, partner forecast and AI business case tied to frontier-lab growth with a downside case 30-40% below reported figures, before your next board meeting.

  3. Write down your own ARR definition (gross or net of partner payouts, single month or trailing three months) and reconcile it to recognized revenue in the next board package.

Agents Are Arriving as Customers Your Seat Pricing Can't Bill

A commerce coalition, a payments race and tools rebuilt for machine traffic all lead to the same P&L question: what does an agent pay for?

The load arrived before the business model

GitHub had to rebuild its Git infrastructure because agent-driven development generates heavy concurrent reads and writes. After separating storage from compute, it reports up to 35x higher write throughput, a figure it measured itself. Cloudflare rebuilt its command-line tool for the same reason. The new cf CLI exposes more than 3,000 API operations and returns machine-readable JSON by default. Both companies redesigned products for a customer that never looks at a screen. The 48% figure covers one week of one developer tool, used by the people most likely to adopt agents first. Read it as a leading indicator, not a market average.

The technical fix is the easy part. Per-seat pricing assumes one human per seat. An agent making thousands of calls an hour under one employee's login breaks your pricing well before it breaks your database.

Who is writing the terms

The Personal Agent Protocol coalition covers the whole transaction. Meta brings consumer agents, Sierra and Genesys bring customer service, Shopify and Walmart bring storefronts, and Stripe brings payment. The protocol's stated priority is giving businesses visibility and control over what agents do. OpenAI, Google, Anthropic, Microsoft, Amazon and Apple are all missing from the partner list. The coalition is either an early land grab that the giants will later join, or the opening move in a fight over standards.

Payments show the same split. Stripe is building outward from settlement with a set of acquisitions: Bridge for stablecoins, Privy for wallets, Metronome for usage billing and OpenRouter for model routing. Cloudflare instead lets agents pay per request at the network edge, before traffic reaches your servers. Someone will set the toll on machine transactions. If it is not you, it will be your payment provider.

The identity problem underneath

Agents need identities of their own. Artificial Ignorance's Charlie Guo reports that delegation, giving each agent its own service-account credentials, now works best for about 90% of tasks. The fine-grained permissions delegation needs do not exist yet. One example is "read all email, but send only to approved recipients." Until they do, an agent that logs in as an employee gets everything that employee can reach. Microsoft's Execution Containers, now generally available, limit what agents can touch. Nothing announced so far settles how agents get billed.

The commercial upside goes to products that agents act through, not just read from. Valon says it signed more than $200M of ARR in six months after opening its mortgage-servicing system to outside servicers. Its agents complete work instead of flagging it for a human. That is a company claim, and signed revenue is not live revenue.

Agents will reach your product whether you price them or not; the only choice is whether you bill them or absorb them.

The smart move

Measure how much of your usage comes from agents first, because you cannot price what you cannot see. Then decide how to license an agent: as a new account class, as part of a human seat, or metered per request. Make that call before a payment provider or protocol coalition makes it for you.

What to do

  1. Have product and finance track agent versus human share of usage across your APIs, command-line tools and web surfaces, and report it on the executive dashboard by quarter-end.

  2. Decide this quarter how agents are licensed on your platform (a new account class, part of the human seat, or metered per request), and model gross margin for your top five products under per-request pricing.

  3. Name an executive owner to recommend within 90 days whether to adopt, extend or counter the Personal Agent Protocol, starting with any commerce, service or payments surface.

The bottom line

Every yardstick leaders use to buy AI came from someone with a stake in it: the price sheet, the revenue run rate, the safety assurance, even the seat count that assumes a human behind every login. All of them slipped together. An AI strategy priced on those figures is priced on marketing, and the problem grows as agents become both your suppliers and your customers. Name one internal owner this week to sign off every AI renewal on cost per finished task, vendor governance and agent share of usage, and refuse any commitment that fails that review.