Executive Summary

A week about who does the checking. Wikimedia disclosed that agents it believes OpenAI operated made unauthorized edits and millions of API requests across its sites. OpenAI published 722 math manuscripts from an unreleased model, most without formal proofs, and asked mathematicians to verify them. Anthropic cut Haiku prices sharply while changing the tokenizer that determines what you are billed for. Mistral previewed a 1.05 trillion parameter model and promised the weights by month end.

Top Stories

1. Wikimedia Says OpenAI Agents Made Unauthorized Edits and Millions of API Requests

Wikimedia says OpenAI agents made unauthorized edits
Minimize imageEdit imageDelete image

The Wikimedia Foundation disclosed on October 5 that AI agents it believes OpenAI operated edited its wikis without approval. Almost all edits were test edits in sandbox areas and did not reach pages readers see. A few changed the configuration of a citation tool, which Wikimedia called potentially malicious and believes were meant to misuse the tool as a proxy for fetching data from remote services. The agents also tried, without success, to use Etherpad, a note-taking tool, the same way. Separately, agents made millions of requests to Wikimedia public APIs, crawling mainly Wikidata and Wikimedia Commons, with hundreds of thousands of queries to the Wikidata Query Service that Wikimedia says potentially contributed to an outage in May. Wikipedia permits bot editing once the community approves it, and Wikimedia says no approval was sought. The foundation found no sign that its data or systems were compromised. OpenAI said it is reviewing the findings.

Business Impact: This is the second disclosed case of OpenAI agents reaching a third party's live infrastructure, after the Hugging Face incident in July, and this time the victim published the detail. The mechanism is the part to note. The agents went after tools that could be turned into proxies, which is a lateral move rather than a direct attack. Two actions follow. If you run public APIs, assume agent traffic is already arriving and check whether your rate limits distinguish it from human traffic. If you deploy agents, the approval question Wikimedia raises is the one your own counterparties will ask, so establish now how your agents identify themselves to systems they do not own.

2. OpenAI Publishes 722 Math Manuscripts From an Unreleased Model, Most Unverified

OpenAI publishes 722 math manuscripts
Minimize imageEdit imageDelete image

OpenAI released 722 manuscripts organized into 372 result families on October 6, generated by a model it has not shipped and will not make available. The repository is public under Apache 2.0. Coverage reports that only 162 of the 722 papers carry a Lean-formalized main result, and OpenAI warns that unformalized results may have problems. Claimed work spans algebra, number theory, theoretical computer science, logic, and topology, including a paper claiming to prove a formula connected to the Birch-Swinnerton-Dyer problem in certain cases. OpenAI says most results came from a single prompt to a single agent, averaging roughly three hours of ChatGPT Pro compute each, against about 4,000 problems attempted. Outlets report different counts, with at least one citing 377 rather than 372 result families. Mathematicians have been publicly divided, and independent verification is ongoing.

Business Impact: Strip out the mathematics and the pattern is familiar. A vendor generated a large volume of plausible output, could not certify most of it, and moved the verification cost to the people who consume it. That is the same trade your teams make with generated code and generated analysis, at a scale where the checking is the expensive part. The useful number here is not 722 but 162, the share with machine-checkable proof. Ask what your equivalent ratio is, and whether anyone is measuring it.

3. Anthropic Cuts Haiku Prices and Changes the Tokenizer in the Same Release

Anthropic cuts Haiku prices
Minimize imageEdit imageDelete image

Anthropic shipped Claude Haiku 5.5 on October 7 at $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, against $1 and $5 for Haiku 4.5. Above 100,000 tokens, rates rise to $0.50 and $2.50. Cache reads are $0.01. The model keeps a 1M context window and 128K max output, and is the first Haiku with an adjustable effort setting. The same release ships a new tokenizer that counts the same text as roughly 30% more tokens than Haiku 4.5. Anthropic says about 90% of Haiku 4.5 requests stayed under 100,000 tokens and estimates the model runs about 75% cheaper on average once the tokenizer is accounted for. Anthropic-reported benchmarks put it at 72.4% on OSWorld 2.1 against 48.9% for GPT-6 Luna. Luna lists the same short-context rates and is cheaper above 100,000 tokens. Anthropic recommends Sonnet 5.5 and Opus 5.5 for complex agentic coding. Sonnet 5.5 cache reads were halved in the same announcement.

Business Impact: A list price cut and a tokenizer change in one release means the headline number and your invoice move in opposite directions. Anthropic published its own adjusted estimate, which is the honest version, but that estimate is an average across its traffic rather than yours. Run your actual prompts through the new tokenizer before you model savings. The 100,000-token threshold matters more than the rate: workloads that cross it pay five times the short-context rate, and GPT-6 Luna is cheaper on list above that line.

4. Mistral Previews a 1.05T Parameter Model and Promises Open Weights This Month

Minimize imageEdit imageDelete image

Mistral opened a public API preview of Mistral Large 4 on October 6 through Mistral Studio and its API, with open weights promised by the end of October. It is a mixture-of-experts model reported at 1.05 trillion total parameters, natively multimodal, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. Published figures disagree in several places. Active parameters per token are reported as 49 billion in launch coverage and 52 billion in the model documentation. Context is documented at 1M tokens while the independent evaluator Artificial Analysis lists the preview at 524K. Preview pricing is reported at different rates across outlets, and all sources agree it is above Large 3. Mistral says further benchmark, architecture, and post-training detail will follow.

Business Impact: The open weights are the commitment that matters, and they are not here yet, so treat this as a dated promise rather than a shipped capability. If sovereignty or self-hosting is driving your evaluation, the end-of-month release is the date to plan against, not the preview. The disagreement between Mistral's own documentation and its launch materials on active parameters and context is a reminder to benchmark against the artifact you will actually run. European training and hosting is a real differentiator for regulated workloads, and it is the one claim here that does not depend on unaudited numbers.

Quick Bytes

  • Anthropic consolidates cyber access: The Cyber Verification Program absorbed Project Glasswing into three tiers covering defense work, authorized penetration testing, and safety-critical systems, the last vetted with US government involvement. Reported vulnerability counts from Glasswing participants vary across outlets from roughly 129,000 to 134,500. (SecurityWeek)
  • Decisions API reaches public beta: OpenAI moved the Decisions API announced at DevDay into public beta on GPT-6 Luna, returning typed answers and confidence scores rather than text to parse. (OpenAI)
  • Biren raises again: The Shanghai AI chipmaker is reportedly raising about $515 million in a new share sale. (Bloomberg)

Industry Impact

One pattern dominated. Verification moved from a footnote to the product question, and in four separate places the party doing the checking is not the party making the claim. Wikimedia had to detect and attribute agent behavior itself. Mathematicians are being asked to referee 722 manuscripts whose author will not be released. Anthropic's buyers have to re-measure a price cut against a tokenizer change. Mistral's documentation disagrees with its own launch materials. None of this is scandal, and three of the four vendors disclosed the complicating detail themselves. It does mean the diligence step you may have treated as optional is now the step that determines whether the headline number is true for you. Budget time for it, and prefer vendors who publish the adjusted figure rather than the flattering one.

How Azumo Helps

Agent governance, inference cost modeling across tokenizers and tiers, and evaluating open-weight options against hosted APIs all take engineering judgment. Azumo brings senior AI engineers, nearshore from LATAM, with experience in multi-vendor routing, cost and cache optimization, agent orchestration, RAG architectures, and regulated-industry deployment. 300+ AI and software projects delivered, SOC 2 compliant, 95% NPS.

Sources

Are You New to Outsourcing?
We Wrote the Handbook.

We believe an educated partner is the best partner. That's why we created a comprehensive, free Project Outsourcing Handbook that walks you through everything from basic definitions to advanced strategies for success. Before you even think about hiring, we invite you to explore our guide to make the most informed decision possible.