CodingSoftware & developer tools
HPE says routing coding work to self-hosted open-weight models cut its AI token costs by 30–50%
Agentic coding is pushing AI token bills sharply higher, from GitHub's metered Copilot pricing to Uber's exhausted budget. In a sponsored MIT Technology Review piece, HPE argues for routing routine work to self-hosted open-weight models and tracking cost per merged pull request.
From the MIT Technology Review · AI article of November 2, 2026 and cited sources: Token costs for coding agents are now a real budget problem. [4] [5] [3] [6]
SOURCES[4] GitHub Copilot is moving to usage-based billing… github.blog[5] Uber's 2026 AI Budget Ran Out After Claude Code Took Over… news.designrush.com[3] Uber Exhausts AI Budget as Claude Code Hits 84% (AI Weekly) aiweekly.co
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- MIT Technology Review · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 2, 2026 article published
- Covered
- Nov 4 edition
- Evidence
- 9 sources across 9 sites
- 1Stop counting tokens, start counting what shipstechnologyreview.com
- 2HPE CFO Myers: Private cloud cut our AI token spend…constellationr.com
- 3Uber Exhausts AI Budget as Claude Code Hits 84% (AI Weekly)aiweekly.co
- 4GitHub Copilot is moving to usage-based billing…github.blog
- 5Uber's 2026 AI Budget Ran Out After Claude Code Took Over…news.designrush.com
- 6Gartner press release, June 24, 2026gartner.com
Reach: not measured yet
Token costs for coding agents are now a real budget problem. GitHub replaced request-based Copilot pricing with credits metered per token from June 1, 2026. Uber reportedly used up its 2026 AI budget within four months after Claude Code adoption surged. Gartner predicts AI coding-agent spending will pass the average developer salary by 2028. [4] [5] [3] [6] [7]
HPE's approach sends planning and novel reasoning to frontier APIs. A gateway routes code generation, test writing and repeated refactors to open-weight models on HPE's own hardware. By HPE's own figures, this cut token costs 30–50% across more than 150 internal apps. Its CFO earlier cited 60% compared with an estimated public-cloud bill. HPE has published no routing logic or evaluation results. [1] [2]
The most useful idea in the piece is to measure output rather than tokens, through cost per active developer and cost per merged pull request. Spend is highly concentrated: at HPE, the top 5% of users drive most consumption, so per-seat budgets miss the real cost. The metric can be gamed by splitting work into smaller pull requests, so it should be tracked alongside rework and defect rates. [1]
Evidence that open-weight models are near frontier quality is weaker than it looks. OpenAI stopped reporting SWE-bench Verified after finding flawed tests and contamination. One analysis finds the agent harness alone can shift scores by about 10 points, and the gap to closed models is wider on private benchmarks. [8] [9]
GitHub's billing change and Gartner's forecast show that agent token spend is becoming a budget line leaders must manage. If task routing and outcome metrics hold up on real codebases, teams could cut costs without slowing delivery.
Read the full assessment
That has not yet been shown.
Research is done; here is the dossier.
Executive brief
The story is HPE advertising. MIT Technology Review labels it sponsored content that its editors did not write. HPE says routing routine work to open-weight models on its own hardware cut its token costs by 30–50% (MIT TR / HPE). Two months earlier, HPE's CFO put the figure at 60% (Constellation Research). Neither figure has been checked by anyone outside HPE. The cost pressure behind the piece is well documented, though. GitHub moved Copilot to token-metered billing, Uber reportedly used up its 2026 AI budget early, and Gartner expects AI coding costs to pass developer salaries by 2028. The idea worth keeping is the metric: cost per merged pull request.
What changed and event timeline
Uber rolls out Claude Code
Reports say the share of engineers using it agentically went from 32% in February to 84% by March 2026, helped by an internal usage leaderboard ().
GitHub ends request-based Copilot pricing
From June 1, "premium requests" are replaced by AI Credits charged per input, output and cached token.
Uber's AI budget runs out
Coverage citing The Information says the full 2026 budget was gone within four months. CTO Praveen Neppalli Naga said he spent $1,200 in one two-hour session ().
Gartner forecast
Gartner predicts spending on AI coding agents will pass the average developer salary by 2028 (;).
HPE CFO cites 60% savings
Marie Myers said HPE's private cloud cut token costs 60% compared with what public cloud would have cost ().
HPE's sponsored piece runs
It argues for hybrid gateways and owned capacity, and reports 30–50% savings across more than 150 internal apps ().
Capabilities and access
- Product: HPE Private Cloud AI, on-premises infrastructure for running open-weight models, with a gateway that sends the hardest tasks to frontier APIs (MIT TR / HPE).
- Models: the piece names no specific open-weight or frontier models or versions.
- Copilot today: Pro is $10 a month and Enterprise is $39 per user per month. Each plan includes credits equal to its price.
Read the full section
- Product: HPE Private Cloud AI, on-premises infrastructure for running open-weight models, with a gateway that sends the hardest tasks to frontier APIs (MIT TR / HPE).
- Models: the piece names no specific open-weight or frontier models or versions.
- Copilot today: Pro is $10 a month and Enterprise is $39 per user per month. Each plan includes credits equal to its price. Completions and Next Edit Suggestions do not use credits (GitHub Changelog).
Technical analysis for researchers and developers
- Why agents cost more: an agent plans, writes code, runs tests and retries. HPE gives no method for that figure.
- Routing design: frontier models handle planning and new reasoning. No routing logic or evaluation results are published.
- Benchmark caveats:
Read the full section
- Why agents cost more: an agent plans, writes code, runs tests and retries. HPE estimates one task uses 10–50× the tokens of a single chat prompt. HPE gives no method for that figure.
- Routing design: frontier models handle planning and new reasoning. Self-hosted open-weight models handle code generation, test writing and repeated refactors. A gateway decides by complexity, data sensitivity, cost and latency. No routing logic or evaluation results are published.
- Benchmark caveats:
- OpenAI stopped reporting SWE-bench Verified. Its audit found 59.4% of 138 hard problems had flawed tests or task descriptions, and it cited contamination (OpenAI).
- The agent harness used to run an evaluation can move scores by up to about 10 points (Kimchi).
Claims and evidence
- HPE savings (30–50% in the piece, 60% from the CFO): both come from HPE.
- Usage concentration: the top 5% of HPE users account for most consumption, and the heaviest user consumes about 450× the lightest (MIT TR / HPE).
- Uber: comes from secondhand coverage of The Information's reporting (DesignRush).
Read the full section
- HPE savings (30–50% in the piece, 60% from the CFO): both come from HPE. The CFO's figure is measured against a hypothetical public-cloud bill (Constellation). Neither figure says whether hardware, power or operations costs are included.
- Usage concentration: the top 5% of HPE users account for most consumption, and the heaviest user consumes about 450× the lightest (MIT TR / HPE). This comes from HPE's internal data only.
- Uber: comes from secondhand coverage of The Information's reporting (DesignRush).
- Gartner: about a quarter of tech leaders spend $200–500 per developer per month on tokens, and about 6% spend over $2,000. No sample size is published (IT Brief).
Context and prior work
- Open-weight models are close on public benchmarks.
- Gartner gives similar advice. It recommends matching models to tasks, context engineering, token thresholds and reviewing costly workflows (IT Brief).
- The billing change follows agent workloads.
Read the full section
- Open-weight models are close on public benchmarks. One analysis puts them 4–6 months behind closed frontier models on public benchmarks, and 8–10 months behind on private ones (Kimchi).
- Gartner gives similar advice. It recommends matching models to tasks, context engineering, token thresholds and reviewing costly workflows (IT Brief).
- The billing change follows agent workloads. GitHub said a quick chat question and a multi-hour autonomous session had cost users the same under the old model (GitHub Changelog).
Limitations, safety and contested findings
- Conflict of interest: the source is advertising, and HPE sells the infrastructure it recommends.
- Inconsistent numbers: HPE's own savings figures differ (30–50% vs 60%) and neither has a stated baseline.
- Weak evidence for "near-frontier": the open-weight claim leans on contaminated or harness-sensitive benchmarks.
Read the full section
- Conflict of interest: the source is advertising, and HPE sells the infrastructure it recommends.
- Inconsistent numbers: HPE's own savings figures differ (30–50% vs 60%) and neither has a stated baseline.
- Weak evidence for "near-frontier": the open-weight claim leans on contaminated or harness-sensitive benchmarks. HPE itself says this does not prove the models are equivalent on every codebase.
- Uber outcomes unclear: reports say Uber cannot yet show what the spending did for feature delivery. They also say it now caps each engineer at $1,500 a month per agentic tool (AI Weekly).
- Gaming risk: cost per merged pull request can be gamed by splitting work into smaller PRs. HPE suggests tracking rework and defects alongside it.
Business and practitioner implications
- Measure the baseline first: track cost per active developer and cost per merged PR before changing infrastructure.
- Plan budgets around power users: a small share of users drives most spend, so per-seat budgets will miss it.
- Owning capacity pays only under sustained load. Usage that is low or bursty favors APIs.
Read the full section
- Measure the baseline first: track cost per active developer and cost per merged PR before changing infrastructure.
- Plan budgets around power users: a small share of users drives most spend, so per-seat budgets will miss it.
- Owning capacity pays only under sustained load. Usage that is low or bursty favors APIs.
- Test routing on your own code: check candidate models against your repositories and tests, not public leaderboards.
- Expect metered billing to spread: Copilot's switch suggests other vendors will follow, so per-team budgets and alerts are worth setting up now.
The source trail.
Sources (9)
Stop counting tokens, start counting what ships
Article text retrieved; extracted text may omit tables or interactive elements.
technologyreview.com