The picks

Claude Fable 5.1

Joint first on the re-based Artificial Analysis Intelligence Index v4.3 (53, level with GPT-6 Astra), first on the official Terminal-Bench v4.0 board at 57.9%, and first on Vellum's quality/speed/cost composite at 65%. On Arena text it sits at 1504 inside a 1-6 tie band with the rest of the Anthropic top group. Costs more per Index task than GPT-6 Astra ($7.63 against $3.26), so price the alternative if volume matters.

The Arena text board is tight at the top: on the 2 September snapshot Fable 5 leads at 1507 with a rank spread of 1-6, with Claude Opus 4.6 (high) on 1505, Claude Fable 5.1 (max) on 1504 and Meta's Muse Spark 1.2 (xHigh) on 1499, so treat the lead as first-among-equals and read the exact tie band off the live board. Artificial Analysis has re-based its Intelligence Index twice this month, and on v4.3 Claude Fable 5.1 (max) and GPT-6 Astra (max) share first place on 53, ahead of Claude Opus 5 on 51 and Fable 5 itself on 50, with Astra costing $3.26 per Index task against Fable 5.1's $7.63. Terminal-Bench has moved on too: the official board now runs v4.0, where Fable 5.1 leads on 57.9%, Opus 5 scores 51.8% and Fable 5 44.5%, so the old 83.8% figure from the 2.1 board no longer describes current terminal work.

GPT-6 Astra

GPT-6 Astra tops Artificial Analysis's independent Terminal-Bench v4.0 run at 59.6% on xhigh and 59.1% on max, with Claude Fable 5.1 next at 55.1% and Opus 5 absent from the published top three. Vals.ai has it first on Terminal-Bench 2.1 too, at 87.3% ahead of GPT-5.6 Sol (85.8%), Fable 5.1 (85.0%) and Opus 5 (84.6%). It has no independent SWE-bench Verified run — Vals archived that board as saturated — and it lists at $10/$50 per million tokens, 2.5x GPT-5.6 Sol, so price the high-volume repo-scale batches carefully.

Challenger: Claude Opus 5 — Still the leader on Vals.ai's independent SWE-bench Verified run at 97.00%, against the 96.0% Anthropic self-reports in the Opus 5 system card, on a board Vals has now archived as saturated and never ran GPT-6 Astra on. On Vals' Terminal-Bench 2.1 it sits fourth at 84.6%, behind Astra, GPT-5.6 Sol and Claude Fable 5.1, and it does not appear in the published Terminal-Bench 4.0 top three.

vals.ai's Terminal-Bench 2.1 table now has Opus 5 fourth at 84.64%, behind GPT-6 Astra (87.27%), GPT-5.6 Sol (85.77%) and Claude Fable 5.1 (85.02%), and its own figure drops to 81.27% if you count its server-side fallbacks to Opus 4.8 on refused tasks as failures. The SWE-bench Verified lead is thin as well: DeepSeek V4 Pro (0813) sits second at 96.40% against 97.00%, as an open-weight model costing roughly $0.02 per task against $1.29 for Opus 5. Worth knowing that vals has since dropped SWE-bench Verified from its coding index altogether, calling the benchmark saturated, so lean on terminal and repo-scale evidence rather than that one number.

GLM-5.3

GLM-5.3 (max) tops Artificial Analysis's open-weights table on Intelligence Index v4.3 with 45, one point clear of Kimi K3 (max) on 44, and the weights have been on Hugging Face since 28 August. On cost the two are effectively level — about $2.01 per Index task against K3's $2.00 — so the separator is throughput, with GLM-5.3 measured at roughly 66 output tokens per second against K3's 37. It ships under Z.ai's own glm-5.3 licence rather than a standard open one, and it takes text-only inputs.

GLM-5.3 ships under Z.ai's own glm-5.3 licence rather than MIT or Apache-2.0, and providers above $10bn in revenue need a Z.ai security review before commercial use, so read the LICENSE file in the repo before deploying. The weights landed on Hugging Face on 28 August, about two weeks after the 14 August API launch, and the reign here is dated from 10 September. Artificial Analysis moved to Intelligence Index v4.3 on 7 September, so any v4.2 scores or cost-per-task figures are no longer comparable.

Gemma 4 26B A4B

A 25.2B-total MoE with only 3.8B active parameters, so it decodes at small-model speed, and Google's quantisation-aware training keeps 4-bit quality close to bfloat16. Be specific about which file you pull: Unsloth measured a naive Q4_0 conversion of the QAT checkpoint at 70.2% top-1 against 85.6% for their dynamic UD-Q4_K_XL build at 14.2GB, so prefer that over Google's own 14.4GB q4_0 GGUF, and budget the roughly 15GB of total memory it wants — it loads on a 16GB machine but leaves little room for KV cache, so plan for modest context rather than the full 256K. The pick is about deployability: Apache-2.0 weights, QAT builds sized for laptops, native function calling and image input, rather than raw benchmark score.

Challenger: MiniCPM5-2B — OpenBMB's official GGUF build is now up, tagged for tool-calling and on-device use, and the file sizes are published — 1.56GB at Q4_K_M, 2.68GB at Q8_0, 5.04GB at F16 — so it clears a 16GB budget with room to spare. What is still missing is independent 4-bit retention data and any tool-calling-under-quantisation numbers. On Artificial Analysis's re-based v4.3 index it sits at 13 against this pick's 17, so it is a deployability story rather than a quality one.

Qwen3.8 27B is still the quality leader in this size class — on Artificial Analysis's re-based v4.3 index it scores 34 at xhigh (28 medium, 26 low) against this pick's 17, with the non-reasoning variant at 22 — but the only official quantised build is Qwen's FP8 repo, so a 16GB fit remains unproven, and the third-party 4-bit builds that do exist, such as Unsloth's 23.4GB NVFP4, are GPU-oriented and well over a 16GB budget. The pick's QAT int4 GGUF is 14.2GB in Unsloth's dynamic build, about 200MB smaller than a naive Q4_0 conversion of the same checkpoint, and wants roughly 15GB of total memory; that gap is the whole argument for it. The title stays under review because a verified 4-bit Qwen3.8 build with published file sizes would likely take the slot.

ARK-ASR-3B

ARK-ASR-3B heads the Open ASR Leaderboard's public listing at 4.76 mean WER with an RTFx of 490.98, ahead of MOSS-Transcribe-preview-2B on 4.87, MOSS-Transcribe-Diarize on 5.17 and Cohere Transcribe on 5.42, with per-dataset results filed in the repo and dated 23 June 2026. The 5.04% on the card is the seven-set average that omits TEDLIUM; add TEDLIUM's 2.79% and you get the board figure. It ships Apache-2.0 with 19 languages, but that is not what separates it from MOSS: only the preview-2B is English-only, while MOSS-Transcribe-Diarize is also Apache-2.0 and covers 50+ languages, so ARK's edge there is accuracy rather than licence or reach.

Challenger: microsoft/VibeVoice-ASR-Streaming-7B — Microsoft has now filled in the gaps: the 7B card ships under MIT, and the streaming technical report (arXiv 2609.02812) quotes a real-time factor at or below 0.104 on an A100, with ten languages supported rather than the four originally tagged. The accuracy numbers are still the authors' own — it does not appear on the Open ASR Leaderboard, which ARK-ASR-3B continues to head at 4.76 mean WER. Worth a look if you need streaming speaker-attributed transcription, but you cannot yet compare it like-for-like on WER.

NVIDIA's Parakeet-TDT-0.6B-v3 is still the sensible choice where throughput decides the invoice — 25 European languages, automatic language detection and a 6.34% average WER, so you are trading roughly 1.5 WER points for the speed — but note it ships under CC BY 4.0, not Apache-2.0. ARK's headline is board-verified rather than card-only: the leaderboard records 4.76 mean WER at RTFx 490.98, including TEDLIUM at 2.79%. Do budget for less throughput headroom than that suggests, though: AutoArk's own rerun of the seven public splits on 8x RTX 4090, scored with the leaderboard scorer, lands at 5.13% WER and an overall RTFx of 197.

BreezeBlue Breeze TTS 2 Open Weights

Top open-weight model on the Artificial Analysis provider-voice leaderboard at 1202 Elo, ninth overall and 82 points clear of the next open-weight model, Fish Audio S2 Pro at 1120. BreezeBlue quotes under 40ms time-to-first-audio and a 0.32 RTF on a warmed H100, but those numbers come from the --fast-all path, which wants about 14.4 GiB and a 24GB card; plain eager inference fits in roughly 7.7 GiB, so 12GB is the practical floor. English and Chinese only, and the weights sit under the BreezeBlue Research and Non-Commercial Licence, so commercial self-hosting still needs a separate arrangement.

Challenger: StepFun StepAudio 2.5 TTS (Aug 2026) — Leads the pick 1207 to 1202 on the Artificial Analysis provider-voice board, seventh overall against ninth. It is proprietary and API-only on StepFun's platform at about $85 per 1M characters, though, with no weights published, so it doesn't displace the pick for self-hosters. StepFun's open-weight audio line remains Step-Audio-EditX and Step-Audio-2-mini.

Open-weight but not open-source: the inference code is Apache-2.0, but the weights, derivatives and self-hosted outputs fall under the BreezeBlue Research and Non-Commercial Licence, and a paid breezeblue.ai subscription only covers outputs from the hosted API — not anything you generate locally. That keeps Apache-2.0 Step Audio EditX (1099 Elo on the same board) the pragmatic fallback if you need commercial self-hosting. The lead is also board-specific and language-limited: on Artificial Analysis's controlled-voices arena Breeze ranks third among open-weight models at 1002, behind Mistral's Voxtral TTS at 1010, and the model speaks English and Chinese only.

GPT Image 2

Still top of both Arena boards: #1 on text-to-image at 1382, 50 clear of MAI-Image-2.6 on 1332, and #1 on single-image edit at 1461, 22 clear of MAI-Image-2.6 and Grok Imagine Image 2.0 (low), which are level on 1439. It also leads Artificial Analysis's text-to-image arena at 1,172, 27 ahead of MAI-Image-2.6 and 45 ahead of Reve 2.1, though on that site's editing arena it has slipped to third at 1,117, behind MAI-Image-2.6 (1,129) and MAI-Image-2.6-Flash (1,120). It has held the generation crown since its 21 April launch, on the strength of prompt adherence, photorealism and text rendering.

Challenger: GPT-Image-2.5-Sunburst — Tops Arena on both the text-to-image board (1421 vs 1381 for GPT Image 2) and the single-image edit board (1520 vs 1461), and leads Artificial Analysis's editing arena at 1,167, well clear of MAI-Image-2.6 on 1,124. On Artificial Analysis's text-to-image arena it is second on 1,179, eight behind stablemate GPT-Image-2.5-Flare (1,187) and eight ahead of GPT Image 2 (high) on 1,171. Both Arena scores are still marked preliminary on a few thousand votes, and at $210.70 per 1,000 images there are far cheaper options for editing at volume.

Editing is no longer a GPT Image 2 story: on Arena's single-image edit board it sits third on 1461, behind GPT-Image-2.5-Sunburst (1520) and GPT-Image-2.5-Flare (1491), with Grok Imagine Image 2.0 (low) fourth on 1439 and MAI-Image-2.6 fifth on 1434. Artificial Analysis's editing arena has it fifth on 1,114, behind Sunburst (1,167), Flare (1,145), MAI-Image-2.6 (1,124) and MAI-Image-2.6-Flash (1,117), with Muse Image a point back on 1,113. On price it is now level with the newer OpenAI models rather than uniquely dear at $211 per 1,000 images against $210.70 for both GPT-Image-2.5 variants, though still far above $67 for Nano Banana 2 and $38.90 for MAI-Image-2.6.

Gemini Omni Flash

Google's own material still frames Omni Flash as the video-first, natively multimodal option: the model card covers both the original Omni Flash and Gemini Omni 1.1 Flash, and clips arrive with audio generated in the same pass rather than needing a separate sound job. The differentiator is step-by-step conversational editing, which Google's API docs say is enabled by the Interactions API, chaining turns with previous_interaction_id against gemini-omni-1.1-flash, and the same model is surfaced in the Gemini app, Google Flow and YouTube. Temper both the resolution and the consistency pitch: output defaults to 720p with 1080p and 4K offered as upscales, and Google's known limitations still say holding complete consistency through edits is a challenge, so multi-turn work needs checking shot by shot.

Challenger: Alibaba Wan 3.0 — Now #1 on Artificial Analysis's with-audio text-to-video board on 1242 to the pick's 1238, and it has taken the no-audio board too, 1334 to 1325, so the pick tops neither. Both sit in a ranked range of 1-2 with audio, so read the four points as a tie rather than a lead; Wan 3.0 also leads the video editing board on 1197, with the pick third. Independent same-seed testing now covers two of our criteria: up to 30 seconds in a single job at 480p, 720p or 1080p, strong identity hold for a lone subject, but uncommanded camera cuts and identities bleeding when characters share the frame.

This is a trailing position on most of Artificial Analysis's boards rather than a hold: Wan 3.0 leads text-to-video both with audio (1242 to the pick's 1238) and without (1334 to 1325), the pick is fifth on image-to-video with audio (1179), and it has slipped to third on the video editing board (1121) behind Wan 3.0 (1197) and MiniMax H3 (1129), which is awkward given we sell this pick on multi-turn editing. Its only surviving board lead is image-to-video without audio, 1365 to Wan 3.0's 1362, and that sits inside the margin. On output, Gemini Omni 1.1 Flash defaults to 720p with 1080p and 4K served as upscales and extends in 10-second steps to a 40-second total, so rivals' single-pass 30-second or native-2K numbers are not like-for-like.

Nemotron-3-Embed-8B

NVIDIA's launch materials put Nemotron-3-Embed-8B-BF16 top of the RTEB multilingual board as of 16 July 2026, on a headline 78.46 average NDCG@10 across the 16 public RTEB tasks its model card actually reports, plus 75.45 on MMTEB Retrieval. The card confirms 34 evaluated languages, a 32,768-token maximum sequence length and 4,096-dimension vectors you can slice to 2,048 or 1,024 and re-normalise, with weights under OpenMDW-1.1 and 1B BF16 and NVFP4 checkpoints for cheaper serving. Mind the basis before you quote it: the live RTEB(beta) multilingual board lists 30 tasks over 22 languages, ranks by Borda count of per-task ranks rather than by mean score, and has had its private column temporarily removed since January 2026, so the placing rests on a vendor snapshot of an open-task view.

Challenger: Gemini Embedding 2 — Google's Gemini Embedding 2 is natively multimodal, mapping text, images, video, audio and documents into one vector space; it went to public preview on 10 March 2026 and reached general availability on 22 April 2026 through the Gemini API, Vertex AI and the Gemini Enterprise Agent Platform. Output defaults to 3,072 dimensions and is Matryoshka-truncatable from 128 up, with 768, 1,536 and 3,072 the recommended sizes, over an 8,192-token context, at list pricing of about $0.20 per million text tokens. It remains API-only, so it is not an option if you need weights you can host yourself.

Microsoft's card for harrier-oss-v1-27b claims state-of-the-art on Multilingual MTEB v2 at release, with a 74.3 average across a 94-language family (0.6b 69.0, 270m 66.5), but it gives no task count, no rival scores and no Borda ranking — so treat any head-to-head against Tencent's KaLM-Embedding-Gemma3-12B-2511 as unconfirmed rather than settled. That card dates from late March 2026 and the model landed in Foundry in April, so pull the live MTEB table before repeating either number. Either way an MTEB v2 task mean and our pick's RTEB NDCG@10 answer different questions: one aggregates all task types, the other is retrieval only.

vLLM

The like-for-like Spheron runs on Llama 3.3 70B FP8 on a single H100 (three-engine, March 2026; two-engine, June 2026) put SGLang 1-4% ahead of vLLM on unique-prompt traffic, which Spheron itself calls run-to-run variance. Both ran vLLM v0.18.0 on the legacy model runner: Model Runner V2 only became the default for dense models in v0.25.0 (11 July 2026) and for every model in v0.29.0 (9 September 2026), so neither run reflects what ships by default today. Pick vLLM for breadth, hardware coverage and release discipline rather than for any current throughput number.

Challenger: SGLang — Near-parity overall and faster on agentic/structured workloads, with major production adoption (xAI, AMD, LinkedIn, Cursor).

Less of a two-horse race than it looks: TensorRT-LLM led the same March 2026 Spheron H100 70B run by 8-13% at 1-50 concurrent requests and about 16% at 100, once its engine was compiled. SGLang's edge is prefix-driven rather than general: on Spheron's June 2026 run with an 80% shared 512-token prefix it cut TTFT p50 by 37% at 50 concurrent requests, while on unique prompts the two sat within 1-4% of each other, and enabling vLLM's --enable-prefix-caching narrowed the TTFT gap to roughly 15-18% at 10 concurrent requests, though at 50 Spheron still puts APC-on vLLM at 2,100 effective tok/s against SGLang's 2,550. Measure your own prefix overlap before committing, and re-run on current releases: those runs used vLLM v0.18.0 on the legacy runner, whereas Model Runner V2 has been the default for dense models since v0.25.0 and for every model since v0.29.0 (9 September 2026).