Revolut Reports 90% Lower Bills

If You Only Read One Thing

The cheapest useful agent isn't necessarily the one with the cheapest successful attempt. Revolut's self-hosted coding fleet lowers its reported bill by changing the default and keeping frontier exceptions. Momentic's QA Bench exposes the other half of that choice: a cheaper bug-finder leaves more defects undiscovered. Buying inference capacity and buying confidence in software are becoming different decisions, with different winners.

Revolut Makes GLM the Default and Reports a Tenfold Lower Bill

Revolut has made self-hosted GLM its default coding model while keeping frontier services for exceptions. Its reported inference invoice is a tenth of its earlier Anthropic and OpenAI spending. The interesting result is that the bank changed how it supplies everyday coding work without claiming that one open model wins every task.

November 3's reporting brought attention to David Abitbol's October 13 production account. He describes serving GLM-5.3 to roughly 2,000 engineers on 56 Nvidia B200 GPUs. That gives the deployment something a small team usually lacks: recurring demand across enough people to share an expensive fleet.

Revolut also improved the service before buying more capacity. Abitbol's chart moves from first-token waits above ten seconds in April to roughly 1.6 seconds in August on the same GPU count, while requests increased. Those are operator-reported observations, rather than a controlled comparison isolating each optimization.

The mechanism starts with how much conversation the server reserves room for. A maximum-length context consumes memory that could serve other requests. Reserving less room for ordinary coding jobs can let more people use the fleet before they start queueing. Cache reuse then avoids repeating some input computation. Both choices depend on actual traffic, so a model's advertised context window tells little about economical serving capacity.

This changes the order of a self-hosting decision. Finding weights that fit is only the beginning. A service has to handle the busiest arrival periods, preserve useful context and leave somewhere for difficult jobs to go. Keeping premium exceptions makes the default easier to change: the open model doesn't have to solve the hardest assignment to handle most recurring work.

The low bill still needs a denominator. Abitbol reports about $0.30 per million blended tokens, but doesn't provide a matched, independently audited cost per accepted change that includes operating labor and repairs. September 8's comparison already showed why shared and dedicated local capacity can have opposite economics. Revolut adds a large production deployment, not a universal discount.

For smaller workloads, paying an API provider to absorb idle capacity can remain cheaper. For a busy fleet, tuning context and concurrency can matter more than another reduction in token prices. The claim weakens if premium escalations and human repair rise enough to erase the invoice reduction; completed changes at comparable quality are the number to follow.

QA Bench Finds More Bugs With Mo, but Cheaper Finds With Codex

An agent can produce a convincing bug report while missing nearly everything else wrong with an application. Momentic's QA Bench makes that gap visible: its own Mo agent finds the most labeled defects, yet still misses about four-fifths with an hour per product area.

The published harness covers four running applications, including Gitea and Metabase, with 1,396 human-reviewed bug labels across 20 product areas. Agents explore fresh copies at pinned commits rather than receiving a known issue to repair. That changes the job from fixing a supplied failure to discovering which behavior deserves investigation.

With 60 minutes per product area, Mo matched 19.3% of the reference bugs, against 6.2% for Codex running GPT-6.1 Sol through Playwright MCP, a browser-control interface. Each matched bug cost $2.67 with Mo and $0.80 with that Codex configuration. The comparison therefore offers different choices for spending less on each finding and discovering more of the failures already known to exist.

A cheap finding doesn't price the defect left behind. A broken checkout that nobody reports can cost far more than hundreds of additional model calls. Conversely, a low-risk internal interface may not justify paying more to investigate every corner. The right agent depends partly on what an undiscovered defect would cost, which no general model leaderboard can supply.

The experiment also gives browser access a more concrete test. Keep the app, task scope and time allowance fixed, then change the agent or browser interface and compare the saved findings. That can reveal whether a configuration actually explores more behavior. A longer, more fluent final report can't answer that question.

Momentic designed and graded the benchmark, and its limitations deserve weight. Agents helped draft the reference set, so all competitors may share blind spots. A model judge matches reports to labels; the runs aren't independently replicated, and the time allowance doesn't equalize spending. The results support a shortlist for exploratory testing, not a declaration that Mo makes software safe.

Revolut can retain premium models for exceptional jobs. Bug-finding needs an extra safeguard because an agent often can't recognize the defect it failed to notice. A second pass over a separately chosen user journey tests something that confidence in the first report cannot. A replacement pays off if, on the same saved cases, it finds more distinct, reproducible defects without exceeding the total review budget.

The Contrarian Take

Everyone says: Once an agent has a browser and enough time, it can replace the manual bug bash.

Here's why that's wrong: QA Bench's strongest system leaves most labeled defects unfound. Running the application removes one obstacle to testing; it doesn't tell the agent which unexpected behavior to investigate. Scripted regression checks and exploratory discovery still cover different failure sets. Replacing the former with an open-ended agent can discard known coverage, while replacing the latter assumes a completeness the benchmark directly contradicts. Agents add an investigator; an empty findings list still can't certify a release.

Under the Radar

Claude threads reduce uploads without reducing token charges. Anthropic's November 2 beta lets a client send only the next messages and a previous response ID. The server reconstructs the conversation, but pricing is unchanged. Threads expire after five minutes without a request, so the application still needs its own history and replay path. This can reduce repeated request processing in long loops; it doesn't supply durable memory or qualify for zero-data-retention use. The API contract separates those promises clearly.
Postman acknowledges bundling agent plugins into Desktop. In a November 3 issue response, a Postman representative says the company bundles its CLI and plugins to simplify setup and will clarify opt-in while reviewing hook approval. The underlying report describes changes to global agent configuration. Installing a familiar desktop app can therefore change what an agent loads later. The report doesn't show data theft, and hook installation alone doesn't prove the hook executed.

Quick Takes

Siri AI opens custom app actions to developers. Apple's November 3 preview lets apps expose content and actions through App Intents in the 27.3 OS betas. Skills can describe how to combine actions, and entities aren't limited to existing Spotlight schemas. For agent design, that offers a typed alternative to finding controls on a screen. Preview access makes the integration testable; it doesn't establish dependable completion of multi-app tasks. (Source)

GitBook's incident reaches connected credentials. GitBook says a November 2 intrusion accessed infrastructure credentials and may have exposed integration tokens, Git Sync secrets and visitors' Assistant questions. Its November 3 update revoked customer API tokens and called for third-party secret rotation. Unlike yesterday's WorkOS exposure, these potentially exposed secrets can authorize operations elsewhere. GitBook attributes the attack to AI agents; the notice doesn't quantify how much autonomy they supplied. (Source)

CrowdStrike separates stopping an agent from asking it to stop. Its November 3 AI Kill Switch announcement gives administrators termination, relaunch blocking and network cutoff controls. The release says termination starts at the Gateway, with endpoint termination coming later. That rollout limit matters: a product-wide availability claim doesn't mean every execution path is covered. Independent enforcement can survive a disobedient model, but only where the enforcing component actually controls execution. (Source)

The Thread

Cheaper default execution can make exception handling the largest remaining source of value. Revolut keeps premium inference for work its default model can't handle. Testing is harder because the exception may stay invisible: nobody knows to escalate a checkout path the first agent never tried. That creates a reason to preserve some investigation outside the default agent's own choices. Sampling user journeys independently can reveal where cheap execution merely concentrated attention on familiar work. The resulting evidence can change the routing policy, rather than just approve or reject one output. A fleet becomes more economical when it learns which work deserves expensive investigation, and that learning requires a sample of the work it would otherwise skip.

Predictions

New predictions:

I predict: By November 18, Postman will publicly document a change to the Desktop opt-in or installation behavior for coding-agent plugins, rather than only restating that the components can be uninstalled. Its representative has explicitly committed to reviewing that path. (Confidence: medium; Check by: 2026-11-18)

November 4, 2026 · 03:46 AM ET