Connect with us

AI & Technology

OpenAI Math Results: 372 Proofs, Real Limits

OpenAI mathematics results: 372 proofs, verified in Lean, but hidden model and no prompts. What works, what doesn’t, who should care.

Published

on

OpenAI Math Results: 372 Proofs, Real Limits

OpenAI mathematics just got a massive public dump: 722 manuscripts covering 372 result families, released on GitHub on October 6, 2026. The model that produced them isn’t public. The prompts aren’t public. The average compute per result was about three hours of ChatGPT Pro thinking. Here’s what you can actually do with it, and where it falls short.

What OpenAI Actually Released

OpenAI published the results in a GitHub repository with protocols for paper revisions and citations. It includes formalizations of many proofs in Lean, a programming language that lets a computer check mathematical proofs. It also includes 10 summaries of the model’s reasoning, estimates of compute spent, and stats about how many problems were attempted.

OpenAI mathematics
OpenAI mathematics

The results span algebra, number theory, theoretical computer science, mathematical logic, and topology. According to Scientific American, among the claims are a solution to the four-dimensional Kakeya conjecture, improvements on important computer algorithms, and progress toward the Riemann hypothesis.

This follows OpenAI’s September announcement that it resolved the Navier-Stokes equation, a Millennium Prize problem with a $1 million reward. That earlier result used a 10,000-agent swarm costing millions in compute. The new results, an OpenAI spokesperson told Scientific American, came from almost every problem being solved in response to a single prompt handed to a single AI agent. Some results may have taken multiple attempts.

How To Use The OpenAI Mathematics Repository

Step 1: Go to the GitHub repository. You don’t need an OpenAI account. The papers, Lean formalizations, and compute statistics are public.

Advertisement

Step 2: Check which proofs have Lean formalizations. Lean verification means a computer has validated the proof’s logic. Scientific American reports that many results have already been verified this way and are therefore all but certain to be correct. Start there if you want results you can trust without re-deriving them yourself.

Step 3: Read the 10 reasoning summaries. These are the closest thing to seeing how the model got there. Don’t expect full transcripts. They’re summaries, not step-by-step derivations.

Step 4: Treat unverified proofs as leads, not facts. Mathematicians will take months to parse the full batch and determine whether the proofs contain novel ideas or are mostly mash-ups of existing techniques.

What’s Good Here

The Lean formalizations are the strongest part of this release. A computer-checked proof removes the guesswork. If a result is formalized in Lean, you can build on it without wondering whether the model hallucinated a step.

Advertisement

The scale is real. 372 result families is not a marketing number. It’s a body of work that will take the field months to digest.

The compute disclosure, while partial, gives you a rough cost picture. Three hours of ChatGPT Pro thinking per average result is a concrete anchor.

What’s Weak

No public model. You cannot replicate these results. MIT mathematician Andrew Sutherland told Scientific American: “Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified. We should ask for receipts.”

No prompts. The advisory group recommended publishing the exact prompt and compute time behind each result. OpenAI released only average compute time and some additional statistics. No prompts.

Advertisement

OpenAI says it is not bound by the advisory group’s recommendations, though a spokesperson said the team is taking the guidelines seriously.

The verification gap. Many results are not yet understood by OpenAI’s own mathematicians, according to a spokesperson who spoke to Scientific American. That means the company is releasing proofs it cannot fully explain.

Who Pushed Back And Why

The Institute for Advanced Study in Princeton, New Jersey, hosts the independent Advisory Group on Mathematics and Artificial Intelligence (AGMAI). The group published its first recommendations in late September. It urged AI labs to release mathematical results promptly through established academic channels, disclose model names, prompts, and compute costs, and refrain from treating releases as marketing vehicles.

AGMAI also asked for equitable access to AI models for the global mathematics community. In a statement, the group warned: “The use of proprietary internal models by AI labs to do mathematical research risks creating a two-tier system where labs outrun the rest of the field, effectively alienating the mathematical community from its own discipline.”

Advertisement

The Institute itself said it does not endorse the practice of testing advanced problems on proprietary models. “It is now the case that AI can output mathematical arguments in situations without the human who prompted it being able to understand the arguments, verify them, or take responsibility for them,” the organization said.

NYU mathematician Tristan Buckmaster, who was working on Navier-Stokes, told the New York Times that mathematicians prompting AI models could be providing information that helped the model reach results. “There’s likely to be a bunch of results where they take someone’s work and then take it to completion,” he said.

Terence Tao has criticized OpenAI and other frontier labs for the “insane” pace of AI-generated results.

Not everyone is against the release. Daniel Litt, a mathematician at the University of Toronto, told Scientific American: “If we want to know the answers to these math questions, I see no reason why we should ask the company to keep them secret from us. To me, it’s going to be a good thing for mathematics.”

Advertisement

Alternatives To Consider

If you want AI-assisted mathematics you can actually verify today, look at Lean itself. The Lean community maintains open tools and libraries. You can run Lean locally, check proofs, and contribute without waiting for OpenAI.

If you want open models, Anthropic and other labs have also produced mathematical results, though the field is still processing them. The Verge notes that OpenAI’s release adds to a growing body of results from OpenAI and rival labs like Anthropic.

If you want to follow the ethics debate, read AGMAI’s public recommendations. They are short and specific.

Who Should Use This

Use the OpenAI mathematics GitHub repository if you are a mathematician who wants to scan for leads, a computer scientist interested in Lean formalizations, or a researcher tracking how AI labs handle disclosure. The Lean-verified proofs are the most immediately useful part.

Advertisement

Do not use it if you need reproducible results today. Without the model and prompts, you cannot replicate the findings. If you are a student looking for a clean textbook explanation, this is not that. If you want to build on unverified proofs without checking them, you are taking a risk.

The bigger issue is structural. OpenAI says it is working to release the model responsibly and will fund workshops and conferences around understanding AI-produced major results. Until that happens, the OpenAI mathematics release is a firehose of claims with a partial verification layer. The Lean proofs are the receipts that matter. The rest is a promise.

Frequently Asked Questions

Is the OpenAI mathematics model available to the public?

No. The model that produced the 372 results is an internal frontier model that has not been released. OpenAI says it is working to release it responsibly, but no date has been given.

Can I verify the OpenAI mathematics proofs myself?

Some of them. Many proofs have been formalized in Lean, a programming language that lets a computer check mathematical logic. Scientific American reports these are all but certain to be correct. Proofs without Lean formalizations have not been independently verified.

Advertisement

What did the advisory group recommend that OpenAI did not follow?

AGMAI recommended publishing the exact prompt and compute time behind each result. OpenAI released only average compute time and some additional statistics. No prompts were published. OpenAI says it is not bound by the recommendations.


Related Articles



About the Author

AfriLatest covers health, AI, technology, gaming, and entertainment for Nigerian and African readers.

Advertisement
Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

AI & Technology

ChatGPT vs Gemini Voice: What Actually Works in 2026

ChatGPT vs Gemini voice modes compared by tier, task and privacy cost — what Live Dictation and Gemini Live actually change for users.

Published

on

ChatGPT vs Gemini Voice: What Actually Works in 2026

If the choice between ChatGPT and Gemini comes down to voice, the buying question is narrow: which tool lets you finish a real task without touching a keyboard, on the plan you can actually afford? Everything else — the demo reels, the sounds-human comparisons — is marketing. A ChatGPT vs Gemini decision made on demo audio is a decision made on nothing.

The documentation that matters landed in a short window. OpenAI shipped Live Dictation, which transcribes as you speak instead of waiting for you to stop, and removed the 10-minute recording cap. According to The Verge, it rolled out globally across web, desktop, iOS and Android, on all plans. OpenAI’s own announcement covers the conversational side: GPT-Live, a full-duplex voice model that listens and speaks at the same time.

Two different products. Most coverage blurs them.

ChatGPT vs Gemini
ChatGPT vs Gemini

The Real ChatGPT vs Gemini Question Is Access Tier, Not Hype

ChatGPT’s voice work splits into two engines. Standard Voice is the free default: it converts speech to text, processes it with GPT-4o or GPT-4o mini, then speaks the reply — which is why it lags. Advanced Voice uses natively multimodal models, picks up tone and pace, and responds in real time. Free accounts get a daily preview of Advanced Voice, not unlimited access. CNET’s hands-on review is explicit about that split.

Then there is dictation, which is not voice mode at all. Dictation is a typing replacement: you speak, the transcript lands in the prompt box, you edit it, you send. Voice mode is a conversation. CNET also notes that Gemini Live already handles interruption — you can talk over it and it keeps up — while Claude’s voice mode sits in beta on mobile with on-screen bullet points, and Perplexity’s assistant answers spoken questions and launches apps on command.

Advertisement

That is the honest state of the field. The features exist across four products. The differentiator is what each vendor publishes about limits.

Free Tier: Where The Demonstrable Gains Actually Sit

ChatGPT vs Gemini
ChatGPT vs Gemini

The strongest evidence in the ChatGPT vs Gemini comparison is unglamorous. A free ChatGPT account now gets live dictation with no 10-minute ceiling, a toggleable live transcript, and an editable prompt box before sending. That is a change a user can measure: a 12-minute recorded note used to require chunking into pieces. Now it does not.

CryptoBriefing’s rollout report dates the start to November 18, 2026, and confirms the upgrade is not paywalled. Record mode — the meeting and voice-note transcriber — stays in beta and limited to Plus, Enterprise, Edu, Business and Pro workspaces. So the free tier gets faster transcription but not the workspace-grade recorder.

For a Nigerian user on a metered connection, that distinction carries weight. Live dictation streams audio as you speak rather than uploading a finished clip, and continuous streaming is not the cheap option on prepaid data. The practical test is a three-minute voice note on your own SIM, not a review written from fibre.

Gemini’s free positioning cannot be judged to the same standard from the available material. CNET documents Gemini Live’s conversational behaviour. It does not document a Google equivalent to the editable live transcript, or a published dictation time limit. Readers should treat that gap as unverified, not as proof of absence.

Advertisement

Paid Tier: GPT-Live, Delegation And What It Does Not Fix

Paid ChatGPT users get GPT-Live, rolled out as GPT-Live-1 and GPT-Live-1 mini, with GPT-5.5 running in the background. When a request needs web search or deeper reasoning, GPT-Live delegates it and keeps talking while the answer is prepared.

OpenAI’s published evaluations put GPT-Live-1 ahead of Advanced Voice Mode on GPQA, BrowseComp and an internal τ³-Voice Telecom variant, and ahead on human preference across matched 5–10 minute conversations. Those are vendor-run tests, one of them on a customised internal variant. They are evidence of direction, not independent proof.

ChatGPT vs Gemini
ChatGPT vs Gemini

What none of it fixes: voice mode runs on the same language model as typed ChatGPT, so it hallucinates the same way. CNET says it plainly — double-check anything important. A confident spoken answer about a drug dosage or a contract clause carries the same risk as a typed one, with less visual friction to catch it.

Privacy And The Cost Of Talking Out Loud

Dictated audio stays tied to the user’s chat history and follows OpenAI’s existing data policies. Translation: a voice memo naming a client, a price or a medical detail becomes a chat record. Anyone who would not type it should not say it.

OpenAI has added SynthID watermarking to supported audio generated through ChatGPT Voice and the API, and opened API access to its provenance verification tool so organisations can check whether an audio file came from OpenAI’s models. That matters in markets where cloned-voice fraud is a live concern — but it covers generated audio, not the transcription of a human speaking.

Advertisement

One disclosure worth flagging: Ziff Davis, CNET’s parent company, filed a lawsuit against OpenAI in April alleging copyright infringement in training and operating its systems. CNET published that alongside its recommendation. Readers can weigh it.

Verdict: Which One To Pick Today

Top pick is ChatGPT, on documented grounds rather than preference. Two independent outlets confirm the same facts: live transcription across web, desktop, iOS and Android; every plan including free; the 10-minute cap gone; an editable transcript before sending. That is a specific, checkable improvement to a specific task.

Runner-up is Gemini Live, and the position comes with a caveat. CNET credits it with genuine interrupt-and-keep-up conversation. But Google’s plan structure, dictation limits and transcript editing are not confirmed in the material reviewed here, so the ChatGPT vs Gemini verdict stays provisional on the Gemini side. Test Gemini Live against your own workload before assuming parity in either direction.

Treat adoption numbers with the same caution. OpenAI says more than 150 million people talk to ChatGPT weekly through Voice and Dictation. CryptoBriefing reports roughly 250 million weekly voice and dictation users by early November 2026, up from over 150 million in July. Two figures that do not reconcile — directional, not precise.

Advertisement

The ChatGPT vs Gemini buying rule stands. Name the version, test one identical task on both, check where the audio goes, and ignore any comparison that opens with how natural the voice sounds. On that test, ChatGPT currently wins the documented case — and a documented win is the only kind worth paying for.

Frequently Asked Questions

Is ChatGPT Live Dictation the same thing as ChatGPT Voice mode?

No. Voice mode is a spoken back-and-forth conversation with the AI. Live Dictation is a typing replacement: you speak, the transcript appears in the prompt box, you edit it before sending. CNET and The Verge both treat them as separate tools serving different jobs.

Do free ChatGPT users get the new dictation features?

Yes, for dictation. Reporting confirms the Live Dictation upgrade, including the removal of the 10-minute recording limit, reaches all plans including the free tier on web, desktop, iOS and Android. Advanced Voice conversations remain a paid feature, with free accounts limited to a daily preview.

How does Gemini Live compare on dictation?

Only partly verifiable. CNET confirms Gemini Live handles interruption well, letting you talk over it. A Google equivalent to the editable live transcript or a published dictation time limit is not confirmed in the available material, so that gap should be treated as unverified rather than as proof Gemini lacks it.

Advertisement



bdde0fd1679cb9f9b710865caff4a9c691f340ee6dca1cb8f32cb44b59614792?s=80&d=mm&r=g

About Xavier Catherine

Xavier Catherine is the Technology Lead at AfriLatest. With a background in computer science and a focus on innovation, he oversees the platform's technology infrastructure. Xavier also contributes reporting on AI, gadgets, and cybersecurity, ensuring AfriLatest remains secure, reliable, and user-friendly.

View all articles by Xavier Catherine
Continue Reading

AI & Technology

AI Data Centres Face Backlash: The 2026 Vote Count

AI data centres lost 40 of 41 local ballot votes in the 2026 US midterms. What the verified record shows on costs, timelines and siting risk.

Published

on

AI Data Centres Face Backlash: The 2026 Vote Count

The tool under review is AI data centres, buildings that draw grid power to train and serve models, and the specification sheet towns are being asked to sign has a hole in it. The benefits arrive as projections. The costs arrive as line items on the electricity bill.

That gap turned into a political liability in the November 2026 US midterms. Of 41 local ballot measures on data centre restrictions across 10 states, voters rejected the industry’s position in every one except Harwood, North Dakota, according to Ballotpedia’s count reported by CNBC. Ohio alone carried 21 of them.

CNN’s voter poll found about a quarter of voters called AI and data centres “crucial” to their vote, behind cost of living and corruption. Asked whether they wanted data centres built in their own community, those same voters opposed it by roughly 3 to 1.

AI data centres
AI data centres

Read the two numbers together and this is not an anti-AI wave. It is a siting fight, and the industry is losing it at the level where siting is actually decided.

What AI Data Centres Actually Deliver, And What They Don’t

The strongest pro-industry argument in the reporting comes from Dan Diorio of the Data Center Coalition, who told CNN that residents of Loudoun County, Virginia, are seeing tax relief from the local data centre boom. He also conceded the industry has to show it better: “People need to see it and feel it.”

Advertisement

No figure accompanied the claim, and no independent look at Loudoun’s rate bill was cited. The same reporting notes Virginia politics have been reshaped by spiking electricity costs tied to data centre demand. A trade group saying residents are better off is a press release. A published tax schedule is evidence.

The clearest verified numbers describe money not yet spent. OpenAI’s 8 gigawatts at the PORTS-Pike Technology Campus in Ohio is the largest planned AI campus in the world, and the project’s own site says phase one construction begins this year with capacity coming online in phases from 2028. The figure of more than $500 billion is a projection, not an invoice. Nvidia is backstopping it, and SoftBank and SB Energy are developing it with the US Department of Energy. None of that is delivered compute.

AI data centres
AI data centres

Why 40 Of 41 Data Centre Votes Went Against The Industry

Kaufman County, southeast of Dallas, hosts two large projects. Forney, its biggest city, is already moving to outlaw new data centres inside city limits. A pro-Talarico super PAC targeted voters within five miles of a proposed or approved project in 55 counties. Talarico carried Kaufman by 1 point and won statewide by 12.

In Ohio, Sherrod Brown beat Jon Husted after branding him “the face of data centers,” a line the National Republican Senatorial Committee itself used in an August memo calling the issue “the anchor hanging around Husted’s neck.” At least 10 Ohio cities passed charter amendments banning large data centres; Granville’s cleared 80 percent.

In South Carolina, Annie Andrews beat Darline Graham after calling her “Data Center Darline,” flipping Colleton County, which Trump won by 19 points. A super PAC aligned with the industry, Building the Future, spent $2.5 million for Graham, and the spending became an attack line of its own.

Advertisement

The Cost Question Is A Pricing Question

Georgia matters because the machinery of control changed without a single vote on a data centre. The state’s Public Service Commission flipped to a 3-2 Democratic majority, giving Democrats control of the body that decides who pays for power and how fast it arrives. Bernstein analyst Madison Rezaei’s read, via Yahoo Finance, is precise: “That makes it a pricing lever on 2030s load, not a stop sign on today’s construction.” About 87 percent of Georgia’s pre-construction credible capacity is dated 2029 or later.

Texas is where the gap between announcement and delivery is measured in gigawatts. Jefferies estimates 13.3 gigawatts of announced developments could be at risk under the moratorium backed by governor-elect Gina Hinojosa, out of 248 planned projects. Texas is projected to supply roughly a quarter of new US capacity through 2030. For AI data centres, the fight has moved from whether to build to who pays for the wires.

AI data centres
AI data centres

What The Backlash Means For Africa And Other Watching Markets

None of the four reports cited here contains a single figure on AI data centres in Africa. That absence is the honest headline, because the capital-mobility claim is documented while the opportunity claim is not.

Alpha Compute CEO Brittany Kaiser told Fox News Digital the results “make us and our investors pause and think if we should be investing somewhere else,” and said the company has already invested in data centres in Canada, Europe and Mexico. That is verifiable: a developer with existing non-US assets describing a reallocation. What is not in the record is any African grid-cost data, any equivalent of a public service commission ruling, or any published tariff structure that would let a reader in Nairobi or Lagos judge whether the same cost-shifting risk applies at home.

The market reaction is the other deliverable. The S&P 500 fell 4.4 percent on Thursday, wiping out about $3 trillion; the Dow dropped 1,280 points and the Nasdaq 4.8 percent. Markets had priced the buildout as certain. Politics repriced it in a day.

Advertisement

Who Should Trust The Buildout, And Who Shouldn’t

Communities with a project in the pipeline should use the lever while they hold it. The Ohio governor-elect’s “conditional moratorium” is the most concrete template in the reporting: developers cover 100 percent of utility costs and build with union labour. It turns an argument about vibes into a contract term, and it is enforceable in a way a protest sign is not.

Investors should treat announced gigawatts as pipeline, not capacity. Georgia’s 87 percent figure is the most useful single number in the material, and Kaiser’s warning about existing builds is the practical constraint: it is hard to pivot after land, power agreements and capital are committed.

Anyone citing the $500 billion Ohio figure as committed money, or the 3-to-1 local opposition as a verdict on AI itself, is overreading both. A quarter of voters called the issue crucial; cost of living ranked higher. On AI data centres the verified record is narrower and more useful than either side claims: the costs land first, the benefits land later, and nobody has published the tax bill yet.

Frequently Asked Questions

Did voters actually reject data centres in the 2026 midterms?

On local ballot measures, yes. Of 41 measures in 10 states tracked by Ballotpedia and reported by CNBC, voters sided against data centres in all but one, in Harwood, North Dakota. Nationally the picture is narrower: only about a quarter of voters called AI and data centres crucial to their vote, but roughly 3 to 1 opposed building them locally.

Advertisement

How much of the announced US data centre capacity is actually operating?

Much less than the headline numbers suggest. OpenAI’s 8 gigawatts at PORTS-Pike in Ohio comes online in phases starting 2028, and about 87 percent of Georgia’s pre-construction credible capacity is dated 2029 or later. The $500 billion figure attached to the Ohio project is a projection, not money already spent.

Does the US data centre backlash affect African markets?

The reporting cited contains no figures on African data centre projects, so any direct impact is unverified. What is documented is that developers are reviewing where they build: Alpha Compute’s CEO said the results make the company pause on future investment and confirmed existing assets in Canada, Europe and Mexico.



f83be4545c83dd77188d9ffa832a9c919c52545cd4732b2e815b8b1ffceb0125?s=80&d=mm&r=g

About Clare Dominic

Clare Dominic is the Editor at AfriLatest. With a background in journalism and a passion for storytelling, she oversees the editorial process to ensure content is accurate, engaging, and relevant. Clare works closely with contributors to maintain the highest standards of journalistic integrity across all news, sports, and lifestyle coverage.

View all articles by Clare Dominic
Continue Reading

AI & Technology

AI Agent Monitoring: What the Telemetry Bill Actually Says

AI agent monitoring now costs more than compute for 69% of enterprises running agents. What Omdia’s data shows, what it misses, and who should buy.

Published

on

AI Agent Monitoring: What the Telemetry Bill Actually Says

AI agent monitoring is the layer that watches what autonomous software does when nobody is looking, and the people selling it now have a number large enough to unsettle a finance director.

An Omdia survey of more than 300 enterprise IT decision-makers across North America and Western Europe, published through TechRadar Pro, found that 69% of organisations actively running agentic AI projects now spend more on observability than on compute and infrastructure combined. The tool built to support AI has become the single largest line item in the AI budget.

That finding is real and worth acting on. It is also a sales document, and the difference between the two matters.

AI agent monitoring
AI agent monitoring

What AI Agent Monitoring Actually Measures

Telemetry is the exhaust of running software: logs, metrics and traces. An agent that reasons, calls a tool, reads a document and writes an answer produces far more of it than a deterministic script, because every intermediate step becomes a signal somebody may want to retain.

The survey found more than half of respondents saw telemetry volume more than triple in the previous 12 months, averaging 3.7x, with AI and machine learning workloads accounting for roughly 43% of that growth. Looking forward, enterprises expect a 9.5x average increase from agentic workloads within two years, and nearly half are bracing for somewhere between 6x and 100x.

Advertisement

“This is just the beginning of the growth curve,” Omdia senior market research analyst Keir Walker said in the research. “We are at the beginning of the hockey stick.”

Using the tool is an architectural decision, not a product choice. A telemetry pipeline sits between the data sources and the expensive backend, routing, filtering, enriching and governing data before ingestion. Omdia names the three capabilities buyers want most: multi-destination routing, data reduction and sampling, and content-based routing.

AI agent monitoring
AI agent monitoring

Why The Telemetry Bill Now Beats The Compute Bill

Average annual observability spend sits at $3.17 million, growing 28% year over year. Thirty-five percent of enterprises report budget increases above 52%, and nearly one in five already spends more than $5 million a year on observability alone. Eighty-one percent are actively hunting for cuts.

The consequence is not only financial. Fifty-nine percent have terminated or delayed at least one agentic AI deployment because monitoring costs were too high. Among the exposure areas hit: cybersecurity monitoring, cited by 27% of respondents, and compliance and fraud detection, cited by 26%.

Read that sequence again. An organisation decides it cannot afford to watch an agent that flags fraud, so the agent is switched off. That is not a technology trade-off. It is a risk decision made by default, by whoever happens to own the observability budget.

Advertisement

Where The Pitch Gets Thinner

Andi Mann, who wrote the TechRadar analysis, is chief product and technology officer at Apica, a vendor that sells data pipelines. The piece runs in TechRadar Pro’s Perspectives channel, which is explicitly a platform for industry contributors. The Omdia data is the evidence; the recommendation to buy a routing layer is the author’s commercial position.

Nothing in the numbers is contradicted by that, but the framing deserves a sceptical read. The same survey shows 22% of enterprises have not begun to consider the data implications of agentic AI at all, while 35% report widespread deployment. That readiness gap suggests a lot of organisations bought agents before they bought a plan.

AI agent monitoring
AI agent monitoring

There is a second check the source material never makes. More agents do not automatically deliver proportionally more capability. Ramez Naam, writing via Noahpinion, argues that parallel agents scale roughly logarithmically or worse for genuine problem-solving gains, even though they work well for raw speed. OpenAI’s Navier-Stokes announcement, widely attributed to around 10,000 agents working concurrently across an 88-hour run, is the case usually cited. The agent count measures how many researchers ran, not what they contributed.

If that holds, enterprises are absorbing exponential telemetry growth to supervise agents whose output is governed by far weaker returns. The bill scales faster than the benefit.

The Data You Cannot Monitor At All

A separate problem sits outside the observability stack entirely. Research from Wake Forest University, presented at the International Conference on Automated Software Engineering in Munich in October 2026, examined third-party skills that extend large language model agents, the plugins that let an agent build a presentation from your notes or scan financial documents for compliance issues.

Advertisement

Ying Zhang, an assistant professor in the Department of Computer Science, and her team tested 17,022 randomly selected skills listed on SkillsMP, described as the largest open-source AI agent skill marketplace, and generated 170,226 outputs. They found credential leakage in two forms: skills that are simply flawed, and skills built to be malicious and push data back to a remote server.

No telemetry pipeline catches that one. It is a supply-chain problem, and it arrives through the same install button your team already uses.

Who Should Buy It, And Who Should Wait

Buy or build the pipeline layer if you are running agentic workloads in production, your observability bill is climbing faster than your compute bill, or you have already paused an agent over monitoring cost. The case for a filtering layer ahead of the backend is strong, and the market reflects it: 97% of respondents have implemented or are actively considering a telemetry pipeline, and 70% are in active evaluation now. Nearly one in four say existing vendor relationships are not a significant factor in the purchase decision, which means the incumbents are not safe.

Wait if your agents are still in pilot. The Omdia sample covers North America and Western Europe only, and the cost thresholds, $3.17 million average spend and $5 million for the top fifth, are calibrated to those budgets. A team in Lagos or Nairobi seeing the same 3.7x telemetry growth is not facing the same absolute figure, but it faces the same volume-based pricing model, with thinner headroom and often harder currency constraints. The number that matters there is the share of cloud spend consumed by logs, not the dollar figure quoted in the research.

Advertisement

Whoever you are, treat AI agent monitoring as a cost-control and risk-control purchase rather than a feature. Organisations that solved it early are, per Omdia, 50% more likely to be prepared for 6x growth and 80% more likely to have dodged the cost crunch. The rest will keep choosing between watching their agents and affording them.

Frequently Asked Questions

What is AI agent monitoring?

It is the observability layer that captures logs, metrics and traces generated by autonomous AI agents as they reason, call tools and take actions. Omdia’s research found 69% of organisations running agentic AI projects now spend more on this observability than on compute and infrastructure combined.

Why do AI agents generate so much telemetry data?

Every intermediate step an agent takes becomes a signal worth recording. More than half of surveyed enterprises saw telemetry volume more than triple in 12 months, averaging 3.7x, with AI and machine learning workloads responsible for roughly 43% of that growth. Enterprises expect a 9.5x average rise from agentic workloads within two years.

Do I need a telemetry pipeline before deploying AI agents?

If agents are in production and your observability bill is growing faster than your compute bill, yes. Omdia found enterprises with a telemetry pipeline are 50% more likely to be prepared for 6x data growth. If agents are still in pilot, the survey data is drawn only from North America and Western Europe, so the cost thresholds may not transfer to other markets.

Advertisement



bdde0fd1679cb9f9b710865caff4a9c691f340ee6dca1cb8f32cb44b59614792?s=80&d=mm&r=g

About Xavier Catherine

Xavier Catherine is the Technology Lead at AfriLatest. With a background in computer science and a focus on innovation, he oversees the platform's technology infrastructure. Xavier also contributes reporting on AI, gadgets, and cybersecurity, ensuring AfriLatest remains secure, reliable, and user-friendly.

View all articles by Xavier Catherine
Continue Reading

Trending