Best in Class, September 2026

Every page in this category picks a tool and sticks with it, because tools have habits worth learning. Models don’t work that way — the leader changes monthly, and half the names circulating in roundup videos are a version behind or slightly wrong.

This is a snapshot dated 8 September 2026, with the leaderboard and the caveat for each task. Treat it the way you’d treat a stock quote.

Transcription link

Two different jobs, two different winners. For live captions — calls, meetings, agents that have to answer while you’re still talking — ElevenLabs’ realtime model runs around 150 ms end-to-end across 90+ languages, with diarisation and word-level stamps.

For accuracy on a finished recording, Microsoft’s MAI-Transcribe-1 (April 2026, the first proprietary speech model out of the MAI group) posts 3.8% average WER on FLEURS across 25 languages and beats Whisper Large v3 on every one of them, in Azure AI Foundry at roughly half the GPU cost.

The name doing the rounds is “Scribe v3”. There isn’t one. The current ElevenLabs speech-to-text line is Scribe v2, with Scribe v2 Realtime as the streaming variant — v3 is their text-to-speech model, which is a different product entirely. Easy to conflate, and it’ll send you to the wrong API docs.

Music, and the licence that matters more than the audio link

Suno v5.5 (March 2026) is the better-sounding model: hand it lyrics and it returns vocals, arrangement, and mix in one pass.

The catch isn’t quality, it’s provenance. In July 2026 the Munich Regional Court ruled that Suno’s models had memorised copyrighted works in training and that those compositions were recognisable in the output — Europe’s first significant judgement that training on copyrighted music needs authorisation. Suno has said it’s building licence-only models to replace the current ones. Until those land, think hard before a Suno track goes into anything commercial with European exposure.

ElevenLabs Music v2 is the answer to that specific problem: trained only on licensed catalogue via Merlin and Kobalt deals, with artists opting in, and every generated track cleared for commercial use. The self-serve tiers cover marketing video, YouTube, podcasts, and ads — film, TV, and Studio Games need Enterprise. Slightly less impressive output, considerably better sleep.

3D link

Meshy is the practical default: clean topology, strong texturing, and exports that land in Blender, Unity, and Unreal without a fight. Its conversational 3D Agent lets you iterate by asking rather than re-rolling, and textures go to 8K.

Rodin beats it on raw geometric detail — it displaces geometry from the texture instead of smoothing everything into a soft blob — at the cost of heavy meshes you’ll retopologise before any engine sees them. Detail-first work: Rodin. Anything headed for a game or a deadline: Meshy.

Roundups still say “Meshy 6”. Meshy 7 shipped in August 2026. Meshy 6 hit stable in January, so the recommendation is right and the version number is two releases stale.

Speech synthesis link

Cartesia’s Sonic-3.6 (August 2026) holds #1 on both Artificial Analysis speech arenas — Controlled Voice at 1,144 Elo, Provider Voice at 1,286 — ahead of Speechify’s Simba 3.2 and Alibaba’s Qwen-Audio-3.0-TTS-Plus. It’s a streaming model, so the ranking comes without the usual latency trade.

For Korean specifically, Eleven v3 is still the more expressive read: audio tags let you direct emotion, pauses, and multi-speaker dialogue rather than nudging sliders. It sits third on Controlled Voice at 1,060 — arena Elo and “sounds right in Korean” are not the same measurement, and this is where that gap shows.

OCR link

PaddleOCR-VL-1.6 (Baidu, May 2026) scores 96.3 on OmniDocBench v1.6 and pulls tables and equations out with their structure intact rather than as flattened text. The headline isn’t the score, it’s the size: roughly 1B parameters, built on ERNIE-4.5-0.3B, beating models hundreds of times larger — and it runs on a CPU, which makes it the rare document model you can put next to the documents instead of shipping them to an API.

Worth pairing with MinerU and Docling if you’re building a pipeline rather than reading one PDF.

Image link

GPT Image 2 (April 2026) took the Text-to-Image Arena at 1,512 Elo — a 241-point lead, the largest margin in that board’s history. The practical difference from every model before it is text inside the image: menus, signs, posters, and labels come out correctly spelled, including dense Korean, Japanese, and Chinese. That single capability retires a lot of “generate the background, add the type in Figma” workflows.

For the free, no-account option that’s still fine for most drafts, see Raphael.

Video link

Text→video: Gemini Omni Flash leads, and generates matching audio — effects, narration, music — synchronised in the same pass rather than as a second job. Google shipped Omni 1.1 Flash in late August and it retook #1 the next day.

Image→video: MiniMax H3 (Hailuo 3.0) is the one people mean, and the “#1” needs qualifying. It leads open-weight image-to-video with audio, and tops the Video Edit Arena at 1,390. On the overall image-to-video board it sits around third, behind Gemini Omni Flash and Wan 3.0. Still the best thing you can run yourself from a single photograph — just not the outright leader the shorthand suggests.

Coding link

The two coding leaders disagree because they measure different things, and that’s the useful part.

Automated benchmarks favour Claude Fable 5.1 (September 2026). One caveat worth carrying: the widely-quoted “95% on SWE-bench Verified” is a third-party leaderboard figure, not a published one — the number Anthropic actually reports, in the system card, is 81.2 on SWE-bench Pro. Different benchmark, much harder, still a lead.

Human votes favour GPT-6 Astra, which took #1 on Code Arena: WebDev on 5 September at 1,797, thirty-five points over Fable 5.1, at pricing that matches Anthropic’s.

Benchmarks reward closing the issue; arena voters reward the app looking right. Pick by which one your work resembles — and note that neither number describes the agent harness you’ll actually be typing into, which is covered in AI Coding Tools.

How to read this page in three months link

Don’t. Re-check the boards — Artificial Analysis for speech, video, and OCR, Arena for image, video, and code — and assume every version number above has moved. What ages well here isn’t the picks, it’s the shape of the questions: realtime or accurate, licensed or merely good, benchmark or ballot, open weights or best overall.

Next link

The same question restricted to models you can download and licence → Open-Weight Picks