Barclays expects half its developers on Claude Code; watch the bill first

*A rollout forecast tells you how many people will use an AI coding tool, and nothing about what it costs or returns.*

Opening read

Barclays has put a date on its AI coding rollout: half its developers on Claude Code, Anthropic's AI coding tool, by year-end. I think the forecast is believable and answers the wrong question. The one large rollout I can point to with a cost detail attached is Uber's, where a conference account says a 2026 budget was gone in four months, without saying which budget. If you are approving a rollout like this, set the spending cap and the measure of success before the headcount.

AI coding: a headcount forecast is a spending forecast

This brief has been tracking whether AI coding tools pay back what companies spend on them, and today a large bank put a date on how many people will use one. Barclays is widening its Anthropic partnership across its global operations, and the announcement carries two forecasts. The first is specific to the coding tool: one in two Barclays developers (50%) on Claude Code before 2026 is out. The second is broader and covers Claude in general, which the bank thinks most of its software engineers will be using by 2027. Anthropic News (https://www.anthropic.com/news/barclays-scales-claude)

The reasonable reading is that this settles whether the tools are ready for cautious buyers. Banks are regulated and slow, so when one expects half its developers on a tool by December, other firms can treat a wide rollout as normal. Demand backs that view. Claude Code's annual recurring revenue (subscription income counted as a yearly rate) stood at $2.5 billion as of February, and Uber gave the tool to about 5,000 engineers in December 2025. Afterward, by an account credited to Uber's chief technology officer in a talk by Cline's Saoud Rizwan, 95% of its engineers used Claude and 70% of committed code (code saved to the shared codebase) came from it. That is the adoption case at its strongest. Latent Space (https://www.latent.space/p/ainews-codex-usage-up-10x-in-6-months) Fireworks AI (https://fireworks.ai/blog/fireworks-nexus) AI Engineer (https://www.youtube.com/watch?v=CoEIs6Xm8m8)

My view departs inside that same Uber account. It also says the company ran through a 2026 budget in four months. The account does not say which budget, so it is no measure of Uber's total spending. Even so, heavy use and an exhausted budget look like one fact seen twice, which makes a forecast of 50% coverage a forecast of spending too. Barclays' announcement supplies the first number without the second. I do not know what Barclays pays, how its contract limits use, or whether code quality held up, and the announcement comes from the vendor. This brief has argued that the return on each seat is hard to predict, and the Uber account suggests the cost of each seat is too. AI Engineer (https://www.youtube.com/watch?v=CoEIs6Xm8m8)

Timing adds a second cost. Spending data from Ramp, a corporate card company, showed OpenAI gaining on Anthropic among U.S. business buyers in August. A bank that wires half its developers into one tool by December is choosing while the alternatives keep moving, and every workflow built around that tool raises the price of changing course. TechCrunch AI (https://techcrunch.com/2026/08/20/openai-is-gaining-on-anthropic-with-business-users-new-data-indicates/)

Here is a check to run before you sign. Take your pilot group and pull three figures for last month: the number of engineers who actually used the tool, the total bill, and the share of committed code that came from the tool. Divide the bill by the active users, multiply by the number of engineers you plan to cover, and count how many months your approved budget lasts at that rate. Counting active users instead of licenses assumes nearly everyone in the wider group will use the tool, which is what the Uber account describes. If the budget runs out before the year does, fix the cap before the rollout grows. Keep the third figure beside the bill as the result you are paying for. For the executive signing off, that reorders the approval: spending cap first, a measurable result second, headcount last.

Model pricing: Google's top model at OpenAI's budget price

Vendors have been cutting model prices for weeks, and this week the cuts reached the top of Google's lineup. Google launched Gemini 4 Argon, which it calls its most powerful model yet, at an introductory $2 per million input tokens and $10 per million output tokens (tokens are the chunks of text a model reads and writes). That matches GPT-6.1 Sol, which OpenAI pitches as near-Astra quality at one-fifth of Astra's price. OpenAI offered Sol after pulling GPT-6.1 Astra over alignment failures, meaning the model did not behave as intended, while GPT-6 Astra itself listed at $10 and $50 as of late September. So the match is between Google's top model and OpenAI's cheaper tier, not between two flagships. Google DeepMind Blog (https://deepmind.google/blog/gemini-4-argon-our-next-era-of-frontier-intelligence/) Don't Worry About the Vase (https://thezvi.substack.com/p/ai-188-gemini-dot-argon) Latent Space (https://www.latent.space/p/ainews-openai-devday-2026-dots-61) Big Technology (https://www.bigtechnology.com/p/what-i-got-wrong-about-metas-muse)

For an engineer picking a default model, that puts a top-tier option at the budget price, with one caution. Google's benchmark claims (scores on standardized tests used to rank AI models) are its own, apart from a first-place ranking it credits to the index run by Vals, an AI benchmarking startup. Anyone setting a 2027 budget should also read "introductory" as notice that the $2 figure can move. TechCrunch AI (https://techcrunch.com/2026/09/30/google-releases-gemini-4-argon-called-its-most-powerful-model-yet/)

AI math: the proof a machine could check

OpenAI's September claim that a large group of AI agents (programs that act on their own) solved Navier-Stokes, a famous open problem about how fluids flow, was disputed. The new turn is a reading from One Useful Thing: the Clay Mathematics Institute, which keeps the Millennium Prize list the problem sits on, seems to regard the result as done, though its official sign-off is still pending. That is a commentator's impression, not a ruling. The question of whether the proof is correct could narrow for a plain reason: OpenAI published it in Lean, a language that lets software check each step. Machine checking does nothing for a separate complaint, the mathematician Buckmaster's allegation that OpenAI knew how far he had gotten on the problem before he made his findings public. The run itself took 88 hours and about 130 billion output tokens, a bill that is far easier to defend when software can confirm what it bought. For an engineer sizing a large multi-agent job, the lesson is to spend at that scale only where a machine can confirm the answer. AI Daily Brief (https://podcasters.spotify.com/pod/show/nlw/episodes/AI-Model-Month-Is-Off-to-a-Blistering-Start-e3ok1na) One Useful Thing (https://www.oneusefulthing.org/p/the-dot-and-the-swarm) OpenAI News (https://openai.com/index/navier-stokes-solution) TechCrunch AI (https://techcrunch.com/2026/09/08/openai-fought-dirty-on-career-making-math-problem-says-nyu-mathematician/) Simon Willison's Weblog (https://simonwillison.net/2026/Sep/8/on-navier-stokes/)

Augur calls

• Call: By 2027-03-31, Barclays or Anthropic will publicly state that Claude Code use at Barclays has reached at least 50% of the bank's developers.
• Criteria: Hit if a press release, results document, or on-record executive statement from Barclays or Anthropic published by the due date reports Claude Code use at or above 50% of Barclays developers. Miss if no such statement exists by then, or the reported share is lower.
• Due: 2027-03-31
• Odds: 30%
• Call: By 2027-01-31, Google's public pricing will list Gemini 4 Argon above $2 per million input tokens or above $10 per million output tokens.
• Criteria: Hit if Google's official pricing page or a Google announcement shows a standard Gemini 4 Argon price above either figure on or before the due date. Miss if both prices are at or below those figures on that date.
• Due: 2027-01-31
• Odds: 35%

What to watch

• Barclays' year-end 50% mark is the next checkpoint. A cost or output figure published with it would say more than the headcount. Anthropic News (https://www.anthropic.com/news/barclays-scales-claude)
• Google labeled its Gemini 4 Argon price introductory. A changed price list would end the price match with OpenAI's Sol. Google DeepMind Blog (https://deepmind.google/blog/gemini-4-argon-our-next-era-of-frontier-intelligence/)
• Clay has not formally accepted the Navier-Stokes proof. A formal ruling would replace a commentator's reading with the institute's own. One Useful Thing (https://www.oneusefulthing.org/p/the-dot-and-the-swarm)
• Google's scores for Argon are mostly self-reported, which is the gap outside testing is meant to close. Anthropic said in September it expects to announce more outside evaluator partnerships in the coming weeks, and who those evaluators are will show how independent that kind of testing can be. Anthropic News (https://www.anthropic.com/news/accenture-embedded-evaluation)

Verification notes

The Barclays figures are forecasts published in Anthropic's own announcement, and the evidence holds no cost or outcome data for the bank. Anthropic News (https://www.anthropic.com/news/barclays-scales-claude) The Uber figures are a secondhand account from a conference talk, and it does not say which budget ran out. AI Engineer (https://www.youtube.com/watch?v=CoEIs6Xm8m8) The GPT-6 Astra list price comes from a September 25 report and may have changed since. Big Technology (https://www.bigtechnology.com/p/what-i-got-wrong-about-metas-muse)

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-10-02?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-10-02)