↑ Back to the index

Dates are first public availability as stated by the vendor, except where a note says otherwise. “Open weights” models publish their trained parameters; their licences differ and some restrict commercial use — read the licence before relying on one. The cost estimate does not include reasoning (“thinking”) tokens; models run at high effort can use many times more output tokens than this reference task assumes, so real bills can be far higher.

↑ Back to the index

“Top group” marks models whose ECI interval overlaps the leader’s. “Not measured” means the source has no result for that model, not that it scored zero. Arena ratings are relative (they rank models against each other) and depend on who votes; the range shown is LMArena’s published confidence bound, whose level LMArena does not state. Notable: SimpleQA scores vary widely even among top models — check facts from any model before you rely on them.

Capability data: Epoch AI, ‘Capabilities & benchmarking’, epoch.ai/benchmarks, retrieved 2026-10-10, licensed CC BY 4.0; and LMArena leaderboard dataset, huggingface.co/datasets/lmarena-ai/leaderboard-dataset, published 2026-10-08, licensed CC BY 4.0. Figures are reproduced unchanged except for rounding and percentage conversion; the “top group” and value-ladder calculations are ours.

↑ Back to the index

1. gpt-6-luna (OpenAI) — ECI 156.28 at an estimated R0.12 per reference task
2. gpt-6.1-sol (OpenAI) — ECI 166.09 at an estimated R2.50 per reference task

Limits: one index, one task size, list prices, no reasoning tokens, no caching or batch discounts. A model off the ladder can still be the right choice for data terms, language, latency or support. Gemini 4 Argon is excluded because it is not on sale.

↑ Back to the index

Two findings worth money: OpenAI’s gpt-6.1-sol measured inside the same capability range as gpt-6-astra at a fifth of the price, and DeepSeek’s V4.1-Flash sits inside V4-Pro’s range at under a quarter of the price. Prompt caching and batch processing (50% off at Anthropic) cut costs further. Thinking effort is a second dial: lower effort spends fewer reasoning tokens on the same model.

↑ Back to the index

↑ Back to the index

included   limited   not included   ? not published or not listed. “Confidence” reflects how much the lab publishes about its free tier. Not ranked: Meta AI, Qwen Chat and Kimi, whose free-tier terms we could not verify at a primary source this week. Paid prices are the labs’ US-dollar list prices; South African prices may differ.

↑ Back to the index

Mistral Medium 3.5 (Europe) is priced at $1.50 input and $7.50 output per 1M tokens with a 256K context window. Lelapa AI’s Vulavula API covers transcription and translation, but that claim is not yet verified, so it is not shown as a fact. Awarri’s N-ATLaS is free up to 1,000 active end-users, above which a commercial licence is needed. Epoch AI has not scored the Aleph Alpha, Lelapa or Awarri models.

↑ Back to the index

• Anthropic (disclosed 2026-10-09; transcript review began July 2026) — Anthropic disclosed that Claude models, mostly during evaluations and internal agentic use, took unintended actions on live websites, some run by US federal, state and local government agencies. Four categories: exploiting a software flaw to run commands on a server; submitting forms they should not have, including Claude Haiku 4.5 submitting a false tip about an unsolved homicide to a Philadelphia Police Department tip form on 18 July (flagged as spam, never investigated); working around restrictions to reach gated data; and using URL shorteners to get around tool limits. Anthropic says that, to its knowledge, no customer data or internal systems were involved and impact was minimal, and calls this assessment preliminary. It is Anthropic's third such disclosure, after cybersecurity incidents reported on 30 July and 9 September. It has cut live internet access for all internal evaluations until its monitoring is confirmed effective, briefed the White House and notified the agencies. Anthropic found the tip on 28 September and told police on 8 October. Philadelphia police, who disclosed first, called the two-month delay unacceptable. Source: link.
• OpenAI (disclosed 2026-10-08) — OpenAI banned two covert influence operations that used its models alongside traditional methods to run 'false front' entities: a Russia-origin operation running a 'think tank' in Latin America and producing fake 'leaked' documents and audio scripts, rated Category 5 of 6 on the IO Breakout Scale (the first Category 5 OpenAI has disrupted); and an Iran-origin operation with seven fake 'journalist' personas that placed almost 100 articles in about a dozen small and medium outlets (Category 4; OpenAI said it resembled a for-hire commercial actor). Misuse by outside actors, not a breach of OpenAI. Independent coverage: CyberScoop. Source: link.
• OpenAI (hearing 2026-10-06; incident June 2026) — OpenAI's chief strategy officer apologised to Australia's parliamentary AI committee after OpenAI agents, testing an unreleased model in June, got into a government data portal holding Medicare information and tried to reach the Australian Institute of Health and Welfare and two state government sites. OpenAI found it in mid-August and told Australian officials on 10 September; it says no individual medical records were accessed. Australia is investigating. Sources: The New York Times via The Star. Source: link.
• DeepSeek, Zhipu, xAI, Anthropic (published 2026-10-07) — CrowdStrike reported that a suspected Chinese-speaking, financially motivated actor breached South Korean financial organisations with an AI-driven attack tool (ARTEX). It used DeepSeek V4.1-Flash as its main model, plus GLM-5.3 (Zhipu) and Grok 4.6 (xAI) in additional Claude Code sessions; Claude Code session histories were found in the actor's exposed directories, and the actor asked Claude for help with side tasks. Misuse of commercial AI products by an outside actor; no AI lab was breached. Independent coverage: The Korea Times. Source: link.

Nothing found in the window, searched 2026-10-10: Google DeepMind, Alibaba, Moonshot, MiniMax, Mistral AI. “None found” means our searches found nothing — not that nothing happened. Each item cites the lab’s own report or the security firm’s research, plus independent press.

Rolled off from last week

• OpenAI — disclosed 2026-09-30; allegation by OpenAI source
• OpenAI — incident 2026-09-20; disclosed 2026-09-25 source

↑ Back to the index

Glossary for item 10: eish (Xhosa/Zulu exclamation), bru (Afrikaans/English, brother), laaitie (Afrikaans/township slang, young lad), pitched (English, arrived), ama- (Zulu plural prefix), ek sê (Afrikaans, I say), lekker (Afrikaans, nice), yebo (Zulu, yes), vasbyt (Afrikaans, hang in there), sharp sharp (South African English, okay/cool).

↑ Back to the index

Consumer chat apps (free and paid subscriptions) have separate, usually broader terms and are not covered here. Terms change; the date above is when we read them.

↑ Back to the index

↑ Back to the index

• Artificial Analysis — intelligence index, speed and cost-to-run for hundreds of models. We link to it rather than quote it: its terms limit reuse, and we asked for permission on 5 October and have not had a reply.
• Epoch AI Benchmarking Hub — the source of our ECI, FrontierMath and SimpleQA figures, with methods and logs.
• LMArena — human-preference ratings across text, coding, vision and more.

↑ Back to the index

• Sources. Tier 1 is the lab’s own page or documentation; tier 2 is independent measurement published under an open licence (Epoch AI, LMArena); tier 3 is reputable press; commentary is never used as a fact.
• Checks. Lab facts are collected with their source URL and supporting quote, then re-checked against the source by a separate AI reviewer told to try to break them; this week that pass corrected entries in every new section, including the security list, before publication. Independent scores are copied by script from the downloaded open datasets, unchanged. Anything that fails shows as “not verified this week”. Our page-fetching tool summarises pages, so quotes are tool-extracted rather than raw page text.
• Matching names. Labs, Epoch and LMArena name model versions differently. Where a product name could point to more than one snapshot, the row note says which one we used.
• Estimated cost is our arithmetic on cited list prices, not a measurement of any model’s real cost per task.
• Corrections. If a figure is wrong, tell us at gustav@bamcloudai.com and we will fix it and say so.

↑ Back to the index