● A free public tool  ·  Built by Poptani Labs  ·  No sign-up  ·  Your data never leaves your browser
AI World · 5-minute read

The week AI got cheaper, smarter — and harder to rank

Cartoon robots and a human hand touching a robot finger against a circuit-board background
Image: 'AI Humans and Robots.jpg' by Bovee and Thill, CC BY 2.0, via Wikimedia Commons

Four stories collided in the last week of October: the workhorse model got much cheaper, the new champion arrived behind a velvet rope, the scoreboard itself was rebuilt — and one of the biggest labs filed to go public.

1. GPT-6.1 Terra: the cheaper workhorse

On October 26, OpenAI released GPT-6.1 Terra, the replacement for GPT-5.6 Terra, built for computer use, visual reasoning and everyday office work. The headline is the price: $1 per million input tokens and $5 per million output tokens — 50% below both its predecessor and GPT-6.1 Sol. It is available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the API as gpt-6.1-terra. OpenAI's own figures claim it scores 15 points higher than its predecessor on Agents' Last Exam (a test of professional computer tasks) at medium effort, for about a third of the cost — and matches the flagship Astra at a tenth of the price per token.

Why it matters: for anyone running AI agents at scale, this is the bill going down. But a cheap per-token price is not the same as a cheap task: Artificial Analysis measured Terra at about $0.40 per task at maximum effort (against $1.40 for its predecessor) and found it verbose — 130 million output tokens across the index versus an 81 million median. One independent analysis found output tokens were 82% of a sample month's bill — the dial that matters is how much the model talks, not just what each word costs.

Sources: gHacks, DataCamp, AIHubMix

2. Gemini 4 Carbon: the champion you can't use

Two days later, Google announced Gemini 4 Carbon, calling it its strongest model, built for cybersecurity and long-horizon work. It went straight to No. 1 on Artificial Analysis's Intelligence Index v5 with a score of 59 — one point ahead of Claude Opus 5.5. Google's own numbers put it at 88.2% on SEC-bench Pro (finding real vulnerabilities in real software), 54.8% on ExploitGym (actually exploiting a reported flaw) and 71% on CWE-bench (turning a finding into a working patch) — evaluated, Google says, without guardrails, which measures raw capability rather than guarded-product performance.

There is one catch, and it is a big one: almost nobody can use it yet. Google is offering Carbon first to a subset of cyber defenders through its Fairwind Program, then to vetted organisations, with wider access promised "as soon as possible" and no date attached. There is no API model ID and it is not in the Gemini app on any plan — though Google has stated an introductory price of $2/$10 per million input/output tokens, rising to $4/$20 later.

Why it matters: the gap between the ranking and the rollout is the story. A launch announcement is not a product release, and the smartest model on the leaderboard is currently a gated preview.

Sources: ByteHerald, Top5Apps

3. The leaderboard moved its own goalposts

On October 27, the day before Carbon's announcement, Artificial Analysis released version 5 of its Intelligence Index — its composite score of model capability, now built from nine evaluations across knowledge work, coding, science and general ability in four equal parts. The big change: 60% of the index now rests on private test sets or held-out answers, so models can't have trained on the test. Scores from the old version do not carry over — Kimi K3.1 Pro fell from 55 on v4.3 to 46 on v5 — and the gap between closed models and the best open-weights model widened from 3 points to 12. Among weights you can actually download today, DeepSeek V4.1 Flash 1017 and GLM-5.3 score highest, at 32.

Why it matters: when the test changes, the ranking changes. A leaderboard is only as honest as its methodology — and v5's shift to private tests is an attempt to keep the scores meaningful as models get better at gaming public benchmarks.

Sources: enjoy.dev, Crypto Briefing

4. Anthropic heads for the stock market

Anthropic filed for a Nasdaq listing under the ticker ANT on October 26 — the lab behind Claude joining the public markets. In the same week, a reference to Claude Mythos 5.5 was spotted on Anthropic's official website, suggesting a release could come as early as next week. That part is unconfirmed: no announcement, no specs, and Anthropic previously restricted the Mythos line to verified cybersecurity and life-sciences researchers. Treat it as a rumour until the company says otherwise.

Why it matters: an IPO filing turns a research lab into a company that reports quarterly — and that changes what it builds, what it charges, and how much it tells the public.

Source: TestingCatalog

The fun bit: the model that trended without existing

Amid the real launches, a model called GPT-6.2 "Bel" did the rounds on social media — discussed, compared and ranked by people who had never touched it. OpenAI hasn't confirmed it exists: the week's most-shared new model may be the one that was never announced at all.

And one genuinely human footnote: on October 29, OpenAI also updated ChatGPT to give shorter answers with less excessive flattery — the chatbot, briefly, was told to stop being quite so impressed with you.

The takeaway

Prices fall, access lags, benchmarks move. Read past the headline number — the price, the ranking, the rumoured name — and ask what you can actually use, what it will actually cost, and who measured it.

← All posts