AI TrackerFrontier desk · AEST

Best model for the job

Picks from benchmarks and price

Text picks use LiveBench, SWE-bench Verified, Aider polyglot and the Hugging Face Open LLM Leaderboard, priced with OpenRouter. Media picks use open public benchmarks (UniGenBench++, GEditBench v2, VBVR-Bench, TTS Arena V2, Open ASR Leaderboard), priced from fal.ai public model pages. LMArena is not used. Every card names its source and dates. Data refreshes daily.

Text and code

Research and search

Pulling evidence, tables, and long context into a usable answer.

Top pick
GPT-6 Astra

Value pick
DeepSeek V4 Flash 0731

Top pick is GPT-6 Astra (listed as gpt-6-astra-max) on livebench Data Analysis at 82.97. Value pick is DeepSeek V4 Flash 0731 at 79.33 on the same benchmark, with a blended OpenRouter price of $0.15 per million tokens. GPT-6 Astra blended price is $40.00 per million tokens. These figures come only from the stored benchmark rows and price snapshots.

  • livebench Data Analysis: gpt-6-astra-max 82.9733 · source
  • livebench Data Analysis: deepseek-v4-flash-0731 79.3273 · source

Prices: OpenRouter public model list.

Pick updated 17 Nov 2026, 08:34 AEST

Reasoning

Multi-step problems, math, and logic that are not arena votes.

Top pick
GPT-6 Astra

Value pick
DeepSeek V4 Flash 0731

Top pick is GPT-6 Astra (listed as gpt-6-astra-max) on livebench Reasoning at 92.65. Value pick is DeepSeek V4 Flash 0731 at 86.63 on the same benchmark, with a blended OpenRouter price of $0.15 per million tokens. GPT-6 Astra blended price is $40.00 per million tokens. These figures come only from the stored benchmark rows and price snapshots.

  • livebench Reasoning: gpt-6-astra-max 92.6538 · source
  • livebench Mathematics: gpt-6-astra-max 96.8062 · source
  • livebench Reasoning: deepseek-v4-flash-0731 86.6345 · source

Prices: OpenRouter public model list.

Pick updated 17 Nov 2026, 08:34 AEST

Coding

Writing and editing code on fixed benchmarks.

Top pick
Claude Sonnet 5.5

Value pick
DeepSeek V4.1 Flash

Top pick is Claude Sonnet 5.5 (listed as claude-sonnet-5-5-max-effort) on livebench Coding at 91.37. Value pick is DeepSeek V4.1 Flash (listed as deepseek-v4.1-flash-max) at 80.04 on the same benchmark, with a blended OpenRouter price of $0.35 per million tokens. Claude Sonnet 5.5 blended price is $8.00 per million tokens. These figures come only from the stored benchmark rows and price snapshots.

  • livebench Coding: claude-sonnet-5-5-max-effort 91.366 · source
  • livebench Coding: deepseek-v4.1-flash-max 80.037 · source

Prices: OpenRouter public model list.

Pick updated 17 Nov 2026, 08:34 AEST

Agentic coding

Resolving real repository tasks and multi-file edits.

Top pick
DeepSeek V4.1 Flash

Value pick
None priced closely enough.

Top pick is DeepSeek V4.1 Flash (listed as deepseek-v4.1-flash-max) on livebench Agentic Coding at 77.27. swebench Verified was not used as the main source: its newest matched model is Claude Opus 4.6 (Feb 2026), behind livebench (Oct 2026). No cheaper model with a matched OpenRouter price scores close enough to name a value pick. DeepSeek V4.1 Flash blended price is $0.35 per million tokens. These figures come only from the stored benchmark rows and price snapshots.

  • livebench Agentic Coding: deepseek-v4.1-flash-max 77.2727 · source

Prices: OpenRouter public model list.

Pick updated 17 Nov 2026, 08:34 AEST

Writing

Instruction following and language work.

Top pick
Gemini 3.8 Flash

Value pick
DeepSeek V4.1 Flash

Top pick is Gemini 3.8 Flash (listed as gemini-3.8-flash-high) on livebench IF at 81.41. Value pick is DeepSeek V4.1 Flash (listed as deepseek-v4.1-flash-max) at 70.03 on the same benchmark, with a blended OpenRouter price of $0.35 per million tokens. Gemini 3.8 Flash blended price is $3.30 per million tokens. These figures come only from the stored benchmark rows and price snapshots.

  • livebench IF: gemini-3.8-flash-high 81.4125 · source
  • livebench Language: gemini-3.8-flash-high 87.7927 · source
  • livebench IF: deepseek-v4.1-flash-max 70.0333 · source

Prices: OpenRouter public model list.

Pick updated 17 Nov 2026, 08:34 AEST

Vision, image, video and voice

Vision and multimodal understanding

Reading images, charts and documents (MMMU-Pro style). Arena votes are not used.

No current, open, machine-readable vision leaderboard exists. The OpenCompass OpenVLM data stops at September 2025 and Epoch's vision data stops in 2025. The current independent source is Artificial Analysis MMMU-Pro, which needs their free API key (ARTIFICIAL_ANALYSIS_API_KEY). Until it is set, no pick is shown.

Pick updated 17 Nov 2026, 08:34 AEST

Text to image

UniGenBench++ semantic consistency score, judged automatically. Price is fal.ai per image.

Top pick
GPT-4o-1.5

Value pick
No cheaper priced model scores close enough.

GPT-4o-1.5 leads UniGenBench++ on overall semantic consistency (0-100) at 95.77 (higher is better). On fal.ai it is token-billed, see page. The top pick is billed in a different unit (see its price link), so no cheaper alternative can be confirmed and no value pick is named.

#ModelScorefal.ai price
1GPT-4o-1.595.77GPT Image 1.5: token-billed, see page
2Nano Banana Pro92.72Nano Banana Pro: $0.15 per image
3GPT-4o92.48n/a
4Imagen-4.0-Ultra-preview-06-0691.65n/a
5FLUX-2-max90.85FLUX.2 [max]: $0.07 per megapixel (one 1024x1024 image)
6Seedream-4-5-25112889.7Seedream 4.5: $0.04 per image
7FLUX-2-flex89.35FLUX.2 [flex]: $0.05 per megapixel (one 1024x1024 image)
8FLUX-2-pro88.35FLUX.2 [pro]: $0.03 per megapixel (one 1024x1024 image)

Source: UniGenBench++ · source data updated 23 Feb 2026 · fetched 17 Nov 2026, 08:34 AEST

Open benchmark (MIT). Scores prompt-following, judged by a model, not image beauty. Models released after the source date are not ranked yet.

Prices are the base rate on each fal.ai model page; size, quality and resolution can change them.

Pick updated 17 Nov 2026, 08:34 AEST

Image editing

GEditBench v2 overall score (automated judge, Elo scale). Price is fal.ai per edit.

Top pick
Nano Banana Pro

Value pick
Seedream 4.5

Nano Banana Pro leads GEditBench v2 on overall score (automated judge, Elo scale) at 1096 (higher is better). On fal.ai it costs $0.15 per image. Seedream 4.5 scores 1089, close to the top, for $0.04 per image.

#ModelScorefal.ai price
1Nano Banana Pro1,096Nano Banana Pro edit: $0.15 per image
2Seedream 4.51,089Seedream 4.5 edit: $0.04 per image
3GPT Image 1.51,071GPT Image 1.5 edit: token-billed, see page
4FLUX.2 [klein] 9B · open weights1,039FLUX.2 [klein] 9B edit: $0.011 per megapixel (one 1024x1024 image)
5Qwen-Image-Edit-2511 · open weights1,038Qwen-Image-Edit-2511: $0.03 per megapixel (one 1024x1024 image)
6FLUX.2 [klein] 4B · open weights1,031FLUX.2 [klein] 4B edit: $0.01 per megapixel (one 1024x1024 image)
7FLUX.2 [dev] Turbo · open weights1,021n/a
8Qwen-Image-Edit-2509 · open weights1,014Qwen-Image-Edit-2509: $0.03 per megapixel (one 1024x1024 image)

Source: GEditBench v2 · source data updated 31 Mar 2026 · fetched 17 Nov 2026, 08:34 AEST

Open image-editing benchmark. Uses the automated Overall column, not the human arena column.

Prices are the base rate on each fal.ai model page; size, quality and resolution can change them.

Pick updated 17 Nov 2026, 08:34 AEST

Image to video

VBVR-Bench overall score: reasoning-heavy image-to-video tasks. Price is fal.ai per second.

Top pick
Seedance 2.0

Value pick
No cheaper priced model scores close enough.

Seedance 2.0 leads VBVR-Bench on overall score (0-1) at 0.5437 (higher is better) among models still available. Sora 2 scores 0.5457 but is excluded: OpenAI shut down the Sora 2 API on 24 Sep 2026. On fal.ai it costs $0.014 per 1000 tokens. The top pick is billed in a different unit (see its price link), so no cheaper alternative can be confirmed and no value pick is named.

#ModelScorefal.ai price
1Sora 2 · excluded: OpenAI shut down the Sora 2 API on 24 Sep 20260.5457Sora 2: $0.10 per second
2Seedance 2.00.5437Seedance 2.0: $0.014 per 1000 tokens
3Veo 3.10.48Veo 3.1: $0.40 per second
4Runway Gen-4 Turbo0.4031n/a
5Wan2.2-I2V-A14B · open weights0.3714Wan 2.2 A14B: $0.08 per second
6Kling 2.60.3691Kling 2.6 Pro: $0.07 per second
7LTX-2 · open weights0.3129LTX-2: $0.06 per second
8CogVideoX1.5-5B-I2V · open weights0.2727n/a

Source: VBVR-Bench · source data updated 15 Apr 2026 · fetched 17 Nov 2026, 08:34 AEST

Image-to-video reasoning benchmark with rule-based scoring. It measures whether the video does the task, not cinematic quality. The authors' own VBVR baselines are excluded.

Prices are the base rate on each fal.ai model page; size, quality and resolution can change them.

Pick updated 17 Nov 2026, 08:34 AEST

Text to speech

TTS Arena V2 blind-vote Elo (a community TTS arena, not LMArena). Price is fal.ai per 1,000 characters.

Top pick
Aurora

Value pick
No cheaper priced model scores close enough.

Aurora leads TTS Arena V2 on blind-vote Elo at 1687 (higher is better). No cheaper model with a published price scores close enough for a value pick.

#ModelScorefal.ai price
1Aurora · ±45, 312 votes1,687n/a
2Inworld TTS MAX · ±27, 1,193 votes1,567n/a
3Inworld TTS 1.5 MAX · ±37, 390 votes1,565n/a
4Hume Octave · ±27, 949 votes1,551n/a
5Inworld TTS · ±22, 1,712 votes1,543n/a
6OpenAudio S2 · ±41, 322 votes1,539n/a
7Eleven Multilingual v2 · ±27, 968 votes1,507ElevenLabs Multilingual v2: $0.10 per 1,000 characters
8Eleven Turbo v2.5 · ±25, 951 votes1,505ElevenLabs Turbo v2.5: $0.05 per 1,000 characters

Source: TTS Arena V2 · source data updated 17 Nov 2026 · fetched 17 Nov 2026, 08:34 AEST

Community blind A/B voting on speech samples (not LMArena). Live leaderboard; preliminary and suspended entries are excluded.

Prices are the base rate on each fal.ai model page; size, quality and resolution can change them.

Pick updated 17 Nov 2026, 08:34 AEST

Speech to text

Hugging Face Open ASR Leaderboard average word error rate. Lower is better.

Top pick
OtterAI/speech-v3.1

Value pick
No cheaper priced model scores close enough.

OtterAI/speech-v3.1 leads Open ASR Leaderboard on average word error rate % (English, short-form) at 3.8844 (lower is better). No cheaper model with a published price scores close enough for a value pick.

#ModelWER (lower better)fal.ai price
1OtterAI/speech-v3.13.8844n/a
2reson8/resonant-1-flash4.0238n/a
3zoom/scribe_v14.15n/a
4microsoft/azure-speech-06-20264.2588n/a
5sophea/asr-k1 (preview)4.6144n/a
6zoom/scribe_v2_pro4.7686n/a
7microsoft/Phi-4-multimodal-instruct · open (mit)5.0213n/a
8ibm-granite/granite-speech-5.0-2b · open (apache-2.0)5.0622n/a

Source: Open ASR Leaderboard · source data updated 20 Oct 2026 · fetched 23 Oct 2026, 08:34 AEST

Hugging Face's open leaderboard, run on public test sets. Lower word error rate is better.

Prices are the base rate on each fal.ai model page; size, quality and resolution can change them.

Pick updated 17 Nov 2026, 08:34 AEST

Voice cloning

Cloning a voice from a short sample.

There is no independent public leaderboard for voice cloning. TTS Arena V2 tests stock voices only and says so. No pick is shown rather than guessing.

Pick updated 17 Nov 2026, 08:34 AEST