A coding tool can become popular before anyone knows what a full rollout will cost. Barclays expects half its developers to use Claude Code by the end of this year. Before you approve a target like that, price the active users in your pilot and decide what result would justify expanding it. This is the Augur dispatch for Friday, October 2nd, 2026. The adoption case is strong. Anthropic says Barclays expects Claude Code to reach half its developers this year, then a majority of software engineers in 2027. Those are forecasts, not completed milestones, and both refer to the coding tool. The uncomfortable comparison comes from Uber. In a conference talk, Klein's Saoud Rizwan relayed an account attributed to Uber's chief technology officer. Very heavy clod use, and a 2026 budget exhausted in four months. The account doesn't identify which budget. It isn't a statement about Uber's total spending, and it doesn't tell us whether the work produced was worth the money. But it gives an executive a sensible reason to ask for the bill before approving the next wave of access. Picture an engineering manager preparing next quarter's request. The pilot is popular, developers want more access, and the easiest number to put on the slide is how many people will get it. That slide can be completely accurate and still leave finance unable to answer the question it was brought into the meeting to answer. The Uber Conference account describes roughly 5,000 engineers getting access in December last year. It says 95% of engineers used Claude and 70% of committed code came from it. Committed code means code saved into the team's shared working record. Those numbers describe adoption and activity. They don't establish that 70% of useful work was automated, or that the resulting software was reliable, or that review became cheaper. Barclays' announcement offers no measured cost or quality result for this coding rollout. It does describe other uses. A knowledge assistant adopted by more than 16,000 colleagues, and a system processing roughly 120,000 emails a day. We shouldn't turn missing coding economics into a claim that Barclays has published nothing about its wider use of AI. Equally, those other workflows don't tell a software director what clawed code will cost per engineer. A regulated bank's adoption is evidence of willingness to deploy. It isn't an audit of the return your own organization will earn, with your systems, your contracts, and your review standards. The first useful check doesn't require a new dashboard. Take last month's pilot bill and divide it by the engineers who actually use the tool, rather than everyone who had a license. Multiply that average by the group you plan to cover. Then, compare the projected monthly bill with the approved budget and the months it needs to last. This is a planning estimate, not a quote. It assumes the wider group will use the tool at roughly the pilot's intensity, and contract discounts or heavier workloads could change the result. Test a busier month too. A spreadsheet is allowed to imagine enthusiasm. Keep the generated share of code beside the bill as an activity measure, but add an outcome that matters. Reviewed changes accepted without extra defects, for example. Imagine a developer getting a first draft finished earlier, only to spend the afternoon correcting it with a reviewer. That could still be a useful tool, but counting the first draft alone would miss part of the work you paid for. Auger's recommendation is a bounded pilot on one repeatable workflow, such as routine maintenance changes. Have the engineering lead own it with a finance partner. Compare with the team's recent baseline, including review effort and defects, and agree on a review after one month. Pause expansion if spending exceeds the agreed allowance or quality worsens. Expand only when the team can show useful, accepted work at a cost it can sustain. Access alone isn't much of an advantage when competitors can buy the same tool. Learning which changes it handles reliably in your code base and improving the review process around those changes could be. Model pricing adds another reason to leave room to change course. Google has announced Gemini 4 Argonne with an introductory price of $2 per million input tokens and $10 per million output tokens. Tokens are the pieces of text a model reads and writes. That matches the standard input and output prices OpenAI publishes for GPT 6.1 Sol. Google is positioning Argonne as a frontier model. OpenAI describes Sol as approaching Astra's capabilities at a lower price. But the discount has a stated destination. Google's own footnote says the prices will become $4 for input and $20 for output after the introductory period. The end date remains unspecified in the announcement. Access matters too. Argon is rolling out to trusted testers and cyber defenders through Google's Fairwind program. The announcement describes broader release as still to come, so an ordinary buyer can't treat it as an immediately available replacement. There is outside testing beyond the Valse ranking. Artificial Analysis has published an evaluation, including a cost comparison that shows why matching token prices doesn't guarantee matching task costs. Different models can use very different amounts of text to finish the same work. Picture the procurement lead putting the introductory rate into next year's budget. The correction belongs in that spreadsheet now, while changing a cell is still the whole job. OpenAI's separate decision to shelve the planned release of GPT 6.1 Astra also needs precise wording. It was a release decision, not a recall of the already available GPT-6 Astra. The mathematics story offers a different test of whether a large computing bill bought something checkable. OpenAI reports that its agents reached a result on the Navier-Stokes problem after 88 hours, using about 130 billion output tokens on that problem. Formalization and verification took another 17 hours. Navier-Stokes concerns how fluids behave. OpenAI released a written proof and a version in lean, a language that lets software check a formal chain of reasoning. The Clay Mathematics Institute has made its own statement. On September 11th, it said the problem had apparently been settled, while pointing to its deliberately unhurried process for evaluating the achievement and assigning credit. That is stronger evidence than a commentator's impression, but it isn't a final prize ruling. The dispute about credit is separate. Mathematician Tristan Buckmaster raised concerns about what OpenAI knew of his work. OpenAI denies seeing the unpublished work and says its investigation ruled out his recent codex prompts influencing the system. Those are competing claims, not something a proof checker can settle. Software can check formal reasoning. People still have to examine whether it proves the intended statement and how the work should be credited. For a leader considering an expensive agent job, decide how you'll judge the answer before authorizing the run. Checkability helps justify an experiment. It doesn't remove the spending limit or turn every technically valid output into useful work. Today's retained auger call concerns Barclays. By March 31, 2027, Barclays or Anthropic will publicly state that Claude code use has reached at least 50% of the bank's developers. Augur assigns that a 30% probability. A hit requires a qualifying press release, results document, or on-record executive statement by the deadline. No such statement, or a reported share below half, is a miss. This tests public confirmation, not private adoption we can't observe. The Google pricing call is withheld for review because Google has already announced higher standard rates. It shouldn't be presented as a fresh, unresolved prediction. Augur calls have dated hit and miss criteria, so predictions can be checked rather than quietly forgotten. They're tracked on the website. Adopt the spending check now. Pilot expansion where reviewed outcomes support it. Watch Argon's wider availability and the formal assessment of the mathematics result without confusing either announcement with a completed decision. The written brief with every source is at augerdispatch.ai. Portions of the Augur Dispatch are AI generated. That is the Augur Dispatch for Friday, October 2nd, 2026. Thank you for listening.