AI Barometer: the best AI for every job right now

Barely a month goes by without an AI laboratory declaring its latest model the smartest ever built. For anyone simply trying to keep up, or trying to pick a tool that drafts a decent e-mail, fixes broken code or turns a prompt into a usable image, the noise is deafening.

This is the first instalment of TechCentral’s AI Barometer, a recurring, evidence-based snapshot of which AI models and platforms lead in the tasks people actually use them for. No vibes and no vendor marketing: every verdict below is drawn from public, independently run benchmarks and leaderboards, each chosen because it tests the use case in question.

The scores were checked in the first week of July 2026 and, given the pace of releases, some will likely have shifted by the time you read this. That churn is precisely why the barometer exists.

How the barometer works

Two kinds of evidence feed this report. The first is arena-style testing: platforms such as Arena (until January known as LMArena, an offshoot of research at the University of California, Berkeley) and Artificial Analysis show users two anonymised outputs for the same prompt and ask which is better. Millions of blind votes are converted into chess-style Elo ratings.

The second is task-based benchmarking, in which models are scored against fixed sets of difficult, realistic problems – resolving real software bugs, answering expert-level science questions or producing the documents and spreadsheets professionals get paid to make.

Both approaches have known weaknesses. Arena votes reward outputs people like, which is not always the same as outputs that are correct. Task benchmarks can be gamed if test questions leak into a model’s training data, which is why the barometer favours contamination-resistant tests. And when leading models sit within a few points of one another, the honest reading is a statistical tie. Where that applies below, we say so.

Everyday assistant and chat

The broadest measure available is Arena’s text leaderboard, which in the first week of July had collected 7.15-million blind votes across 369 models. Anthropic’s newly released Claude Fable 5 tops the board on an Elo of 1 509 – and, remarkably, Anthropic models fill all five top slots.

Google’s Gemini 3.1 Pro Preview is the best of the rest on 1 486, with OpenAI’s GPT-5.5 close behind on 1 481.

The caveat: the top 10 sits within roughly 30 Elo points, so for casual use any of these will serve well. Price separates them more than quality does: on API rates listed by Arena, Gemini 3.1 Pro costs US$2/$12 per million input/output tokens (about R33/R195) against Claude Fable 5’s $10/$50 (about R163/R814), making Gemini the clear value pick.

That comparison applies to businesses building on the models’ APIs, though; for consumers, the flagship subscriptions from OpenAI, Google and Anthropic are all priced within a few dollars of one another, whatever the underlying token costs.

Co-working and office tasks

For AI that does real work – producing the documents, presentations, spreadsheets and analyses of everyday office life – the most relevant public test is GDPval, a dataset of 220 tasks that OpenAI built with industry professionals across 44 occupations. Artificial Analysis runs models through it agentically, with web and shell access, and ranks the deliverables by blind pairwise comparison. Anthropic’s Claude Opus 4.8 leads decisively on an Elo of 1 890, some 121 points clear of GPT-5.5 on 1 769 – a notable result, given that OpenAI created the benchmark.

Coding

The barometer’s coding yardstick is SWE-bench Pro, run as an independent public leaderboard by Scale AI. It asks models to resolve real issues from actively maintained software repositories and was designed specifically to resist training-data contamination, a documented problem with the older SWE-bench Verified test. On the independent board, OpenAI’s GPT-5.4 leads on 59.1%, in a statistical tie with Meta’s newly listed Muse Spark on 55%, with Anthropic’s Claude Opus 4.6 next on 51.9%. One flag for readers weighing vendor claims: Anthropic says its new Claude Fable 5 scored 80.3% on the same task set, but that figure was produced with the company’s own tooling and does not yet appear on the independent leaderboard. Until it does, the verified crown belongs to OpenAI.

Writing

Writing quality is notoriously hard to score, and the most serious public attempt is the independent EQ-Bench Creative Writing benchmark (version 3), which combines rubric marking with Elo-style pairwise judging across 32 varied prompts. Claude Fable 5 leads on an Elo of 2 192, narrowly ahead of Claude Opus 4.7 on 2 179, with GPT-5.5 third on 2 019. An important caveat applies here: the judge is itself an AI model, so results are directional rather than definitive – though the benchmark also publishes the raw outputs, so sceptical readers can judge the prose for themselves.

Research

Research splits into two skills. For deep reasoning, the reference test is Humanity’s Last Exam, 2 500 questions crowdsourced from subject-matter experts precisely because frontier models could not answer them. On the official leaderboard run by Scale AI, Google’s Gemini 3.1 Pro Preview leads on 46.4%, in a statistical tie with OpenAI’s GPT-5.4 Pro on 44.3% – meaning the best models still fail more than half the paper. (Anthropic reports a higher 53.3% for Claude Fable 5, but that score is vendor-reported and not yet on the official board.)

For live web research, the benchmark to watch is BrowseComp, OpenAI’s 1 266-question test of an agent’s ability to hunt down hard-to-find information online. One honesty note: unlike every other test in this report, BrowseComp has no independently run public leaderboard – the laboratories report their own scores. On those self-reported numbers, OpenAI’s GPT-5.5 Pro is top on 90.1%. In short: Google for depth; ChatGPT, on its own numbers, for tracking things down online.

Image generation

Image models are ranked almost entirely by blind human preference, and the Artificial Analysis Image Arena is the busiest public board. OpenAI’s GPT Image 2 (high) leads on an Elo of 1 339 across more than 13 000 comparisons – 58 points clear of second-placed Reve 2.0 (1 281), with Microsoft’s MAI-Image-2.5 third on 1 271.

Video generation

Video is the category where conventional wisdom is most out of date. On the Artificial Analysis Video Arena’s text-to-video board (with audio), ByteDance’s Dreamina Seedance 2.0 leads on 1 224, followed by Alibaba’s Wan 2.7 (1 160) and HappyHorse 1.1 (1 154). Google’s Veo 3.1, the strongest Western entrant, sits only 10th on 1 095.

On the silent board, Alibaba’s HappyHorse 1.0 leads on 1 290. The top four slots on the board all belong to Chinese developers, and the costs remain steep: producing a minute of 1080p footage runs from about $9 to $24 (roughly R145 to R390) on the model creators’ own APIs.

Music generation

Music benchmarking is younger and thinner than any category above, so treat these results as directional. On the Artificial Analysis Music Arena, which uses the same blind-vote method as its image and video boards, Suno v5.5 tops both leaderboards – instrumental (Elo 1 194) and vocals (1 168) – ahead of Mureka V8 in both, with Suno’s earlier V5 holding third place on each board.

The scoreboard

* Statistical tie. † Lab-reported; BrowseComp has no independently run public leaderboard. AA = Artificial Analysis. All figures checked in the first week of July 2026

Until next time

A barometer measures pressure at a moment in time, and the pressure in this industry rarely holds steady for a quarter. Several of the results above are statistical ties that a month of new releases could flip, and at least two categories – video and music – are moving faster than the benchmarks that measure them. We’ll check back with another reading in two or three months to see what’s changed.  – © 2026 NewsCentral Media

• Subscribe to TechCentral’s daily newsletter
• Get breaking news alerts on WhatsApp