GenAI Daily - October 25, 2026: TypeSafe Jev Data Leak, OpenAI Decisions API Goes GA, Replit's $100M Provider Switch
Top Stories
TypeSafe AI Discloses Leak of Every Customer's Jev Output
TypeSafe AI disclosed on October 24 that a leak exposed all customer Jev model outputs. CEO Diogo Almeida posted the news himself on X.
The exposed data is Jev's "noul" output type, a single probability between 0 and 1 that answers a yes-or-no question.
Each noul records a decision a customer's software made, often about a real person or transaction.
Developer Wes Bos confirmed his own output was among those exposed.
Jev's adoption makes the blast radius large.
The model launched about a month ago and spread quickly through coding agents, fraud filters and classification pipelines.
Vercel said 35.3% of its paying AI Gateway customers adopted Jev within two weeks of launch, a platform record.
TypeSafe's docs say enterprise customers get zero data retention.
They also say customer requests are not used for training.
It is not yet clear how those commitments apply to this incident. Separately,
UpGuard found 24 deliberately configured lookalike domains with working mail infrastructure
tied to the Jev and TypeSafe names, so expect phishing aimed at affected teams.
Why it matters: Teams that wired a fast, cheap decision model into fraud, moderation or routing paths should now audit what those logged outputs reveal and add vendor-side data-handling checks to procurement.
OpenAI's Decisions API Reaches General Availability at $0.10 per Million Input Tokens
OpenAI made the Decisions API generally available on October 22, priced at $0.10 per million input tokens.
Requests go to a dedicated POST /v1/decisions endpoint, and gpt-6-luna is the only supported model for now.
The endpoint returns typed answers (a probability, a choice or a score) in place of text.
OpenAI says it is up to 10x faster than running the same model through the Responses API.
Cached input costs $0.01 per million and output tokens are free.
Cache writes are billed, and regional processing premiums and long-context multipliers still apply.
The GA release also
raised the limit to 400 questions per request, up from 200.
It adds prompt caching and native GPT-Live integration.
Zero Data Retention and HIPAA use are open to eligible customers.
The architecture differs from TypeSafe's Jev.
TechCrunch reports OpenAI took ChatGPT Luna and reconfigured it for faster inference at lower cost, rather than building a purpose-calibrated decision model.
Neither vendor has published head-to-head accuracy figures.
Why it matters: Teams now have a major-lab option for classification and routing at Jev-class pricing, and with the Jev leak fresh, a ZDR-eligible option from a larger vendor is easier to justify, though accuracy on your own data still has to be tested.
Replit UPDATE: OpenAI Reportedly Won a $100M+ Annual Contract With 30%+ Discounts
The Information reports that
OpenAI gave Replit more than 30% off to win a contract worth over $100M a year, after Anthropic refused to extend Replit's prior discount without a higher spending commitment.
That adds commercial detail to last week's dispute over CEO Amjad Masad's remarks.
A person familiar with the deal told TechCrunch the Anthropic partnership is still in place, and OpenAI is now the default provider for Replit Agent while customers can still pick Claude.
The report points to a wider pattern of discount-driven switching.
At Rippling, OpenAI's share of coding work grew from 16% to 52% while Anthropic's fell from 57% to 18%, and OpenAI offered DataCamp 40% off for Codex.
Vercel's gateway data shows Anthropic's overall spend share fell from 63.7% to 43.9% in September.
OpenAI projects its compute spending this year will exceed its revenue.
AI Weekly, citing The Information
Why it matters: Large coding-agent buyers have leverage on price right now, so teams negotiating model contracts should expect competing discount offers and should keep their routing layer provider-agnostic.
Key Developments
Study: Coding-Agent Harness Choice Can Triple Cost Without Improving Accuracy
Shandong University researchers ran five models through Claude Code, OpenCode and mini-SWE-agent.
On the larger task set, the most elaborate and the most minimalist harnesses produced statistically equivalent success rates within a five-point margin, though a third harness performed worse.
Cost per task varied by as much as threefold across harnesses using the same model.
Fixed system prompts and tool schemas drove the bill: 16,581 tokens for Claude Code versus 7,025 for OpenCode and 829 for mini-SWE-agent, resent on every step.
The paper also finds that single-run benchmark gaps are noisy.
Rerunning identical SWE-bench Verified setups moved scores by 2.2 to 6.0 percentage points, wider than most harness-to-harness differences the paper measured.
The smallest harness was not always cheapest: on DeepSeek-V4-Flash, mini-SWE-agent made more calls than Claude Code and cost slightly more.
Impact: Evaluate coding agents on cost per solved task for your exact model and harness pairing, with repeated runs, rather than on a vendor's leaderboard delta.
Gartner: 80% of Neoclouds Will Lose Independence by 2030
Gartner predicts more than 80% of neocloud providers will cease independent operations by 2030 as AI infrastructure spending runs ahead of revenue.
Gartner says data center owners would need $8.5 trillion in cumulative revenue to cover announced spending and a 12% return, while attributable AI revenue will reach no more than $3.6 trillion.
The firm did not say how many it expects to be acquired or merged versus shut down.
The debt picture is worsening the pressure.
24/7 Wall St. reports CoreWeave pays $640M a quarter in interest against $128M in operating income.
Impact: Teams signing multi-year GPU capacity contracts with neoclouds should check counterparty financial health and add assignment and continuity clauses, since acquisition is the more likely outcome than shutdown.
Linux Kernel Drops Mandatory AI Disclosure Tag
The Maintainers Summit on October 8 in Prague reached consensus to deprecate the "Assisted-by" tag and remove the documentation requiring it.
Critics called the tag unenforceable, noting that AI-written commit messages are already hard to tell from verbose human ones.
Torvalds says competent developers should use AI as a tool but wants to discourage "drive-by patches" from people who don't understand what they submit.
Accountability stays with the human submitter, and review standards are unchanged.
Impact: Internal policies that rely on self-reported AI authorship face the same enforcement problem, so many teams may shift toward review depth and contributor accountability.
Product Launches
Gemini Enterprise Single-User Quota Limits (GA)
Gemini Enterprise now offers generally available single-user quota limits, letting administrators set per-user usage caps by edition and location to protect shared subscription quota pools.
This gives admins a direct control against one heavy user draining a shared pool, a cost problem other enterprises are hitting with coding agents.
Releasebot Google Release Notes
Why it matters: Per-user caps let administrators keep shared subscription quota pools predictable and prevent a single heavy user from exhausting capacity for everyone else on the team.
Gemini Enterprise Agent Platform CodeMender Updates
CodeMender updates on October 21 enable deep scans by default, broaden file discovery and add a health-check command.
Teams running it for automated code-security fixes get more coverage without changing configuration.
Releasebot Google Release Notes
Why it matters: Deep scans on by default and broader file discovery mean security teams get wider automated coverage out of the box, and the new health-check command makes it easier to confirm the tool is working as intended.
Funding & Deals
Modal Raises $750M Series D
New York-based Modal runs serverless GPU and CPU infrastructure for inference, training and agent sandboxes, and was valued at $15.75B post-money.
It has crossed $1 billion in annualized run rate, and sandboxes alone are roughly a $300 million run-rate product.
Usage has tripled since its Series C five months ago.
Led by Accel.
Why it matters: The round shows investors are backing infrastructure for inference and agent workloads at scale, with sandboxes emerging as a substantial revenue line.
Pathos AI Raises $237M Series E
New York-based Pathos AI applies AI to drug development and raised at a post-money valuation of around $3B. Led by Bristol-Myers Squibb. The round places a large pharmaceutical company at the head of an AI drug-development financing.
Why it matters: A major pharma company leading the round signals continued industry appetite for AI-driven drug development.
Albert Invent Raises $55M Series B
Albert Invent sells AI tooling for materials and chemical R&D. The round was one of the larger AI deals of the week of October 18. Led by TCV.
Why it matters: The raise shows growing investor interest in applying AI to materials and chemical research and development, beyond software and consumer use cases.
Tomorrow's Watch List
- Follow-up from TypeSafe on the Jev leak: the root cause, whether zero-retention enterprise accounts were affected, and customer notifications.
- Any response from Anthropic on enterprise discounting after the Replit and Rippling reports. Its public S-1 is also still expected.
Related reading: Check out this week's [Deep Insights analysis] for strategic context on these developments.
GenAI Daily is curated by rAInvent. For in-depth analysis, see the rAInvent articles.
