This was the week Google undercut OpenAI on image prices, Kuaishou announced a video model you can mostly not use yet, and voice models stopped being a side project. Here is what happened in the week to October 13, 2026, plus the voice releases from the three weeks before it that belong in the same picture.
The headline: image rankings and image prices have split apart
GPT Image 2.5 (September 8) and Nano Banana 2.1 (October 6) now sit at opposite ends of the same trade-off.
| GPT Image 2.5 Sunburst | GPT Image 2.5 Flare | Nano Banana 2.1 | |
|---|---|---|---|
| Arena text-to-image (Elo) | 1425 (#1, preliminary) | 1398 (#2, preliminary) | 1328 (#6, preliminary) |
| Artificial Analysis price per 1,000 images | $210.7 (max quality) | $210.7 (max quality) | $33.6 |
| Token pricing (image output, per 1M) | $30 | $30 | $30 ($15 Batch) |
| Free access | None in the API | None in the API | None in the API; reportedly paid plans only in the Gemini app |
OpenAI took the top two Arena spots without touching token prices, but what you pay per image still depends on the quality setting — about $0.21 at max. Google’s update is the mirror image: it ranks sixth, yet at $0.0336 per 1K image it costs about a sixth as much on Artificial Analysis’s test and adds search grounding and up to 14 reference images. Both sets of Arena scores are still preliminary, with about 17,000 votes each for the GPT Image 2.5 entries and 5,300 for Nano Banana 2.1.
One correction worth making: Nano Banana 2 is Gemini 3.1 Flash Image, not Gemini 3 Pro Image as our page previously said. We have fixed it. Google lists 2.1 as the recommended replacement but has announced no shutdown date for the older model on the Gemini API.
We updated the Best AI Image Generator ranking: GPT Image 2.5 replaces GPT Image 2 and enters at #1 as “Best overall” (Midjourney v8.1 moves to #2), and Nano Banana 2.1 replaces Nano Banana 2 as “Best value” — not “Best free”, because it has no free API tier.
Video: Kling 4.0 is announced, not launched
Kling 4.0 was announced on September 28–30 with native 30-second clips, up to 10 keyframes, up to 15 combined reference assets and two-channel stereo audio. What you can actually use is Kling 4.0 Flash, in limited early access that Kling’s announcement ties to its Ultra yearly plan. The full model is due “in October” with no exact date, and Kling’s own API pricing page says its current packages don’t support Kling 4.0.
Meanwhile Kling 3.0 is the version with an API and a price list: $0.084 per second at 720p and $0.112 at 1080p without native audio, $0.126 and $0.168 with it, and $0.42 at 4K. We corrected our page, which had quoted a third-party ”~$0.067–0.17” range, and added Kling 4.0 to the “On the horizon” list. We then re-ran the whole Best AI Video Generator ranking against Artificial Analysis’s blind-vote boards: MiniMax H3 is now “Best overall”, Wan 3.0 leads text-to-video, Seedance 2.5 and Gemini Omni 1.1 Flash enter the six, and Seedance 2.0 and Veo 3.1 fall to #7 and #19 on text-to-video.
Voice models: five launches in three weeks
| Model | Date | Who | Price signal | What to know |
|---|---|---|---|---|
| Gemini 3.8 Flash TTS | Sep 23 | $0.00225 per 10 s of audio through 2026; doubles on Jan 1, 2027 | Voice design from a text prompt, 130 languages; older TTS previews shut down Nov 17 | |
| Qwen-Audio 3.1 | Sep 23 | Alibaba | List prices cut about 70% (TTS), 85% (Realtime), up to 95% (ASR) — vendor claims | Five models including a full-duplex Realtime model; API-only |
| Eleven v4 | Sep 28 | ElevenLabs | $0.08 per 1,000 characters ($0.04 for Turbo) | #1 and #2 on Artificial Analysis’s Speech Arena; 90+ languages; audio tags |
| MAI-Voice-2.1 | Oct 1 | Microsoft | $22 / $15 per 1M characters (Flash) | 23 languages; public preview with no SLA |
| Nova 2.5 Sonic | Oct 5 | Amazon | Same price as Nova 2 Sonic | Speech-to-speech for voice agents; 18 voices in seven languages, 256K context |
Prices are quoted in different units — tokens, characters and seconds — so they cannot be compared one-to-one without your own usage numbers.
We also published a new Best AI Voice Generator ranking, built on Artificial Analysis’s blind-vote Speech Arena. Eleven v4 holds #1 and #2 — Turbo at Elo 1330 and v4 at 1326, a statistical tie — while Qwen-Audio 3.1’s TTS-Plus tier is #3 at $19.30 per 1M characters and Gemini 3.8 Flash TTS is #5 at $16.50: between a fifth and a quarter of Eleven v4’s $80, for scores 30–60 points behind. Two caveats from our research: the ranked Qwen entry is its character-billed TTS-Plus tier, not the TTS-Flash model whose token pricing we list, and on the cloned-voice board Gemini drops to #17.
Language models this week
- Mistral Large 4 (October 6) entered public preview: 1.05T parameters with 52B active, 1M context and image input, at $1.36 / $4.18 per million tokens — half price during the preview. Weights are due at the end of October, so it is still listed as not open-weight until they ship. See our full story.
- Claude Haiku 5.5 (October 7) completed Anthropic’s 5.5 family at $0.10 / $0.50, about 75% cheaper than Haiku 4.5. More in Anthropic completes Claude 5.5.
- GPT-6 reached ChatGPT on October 7 with a new Intelligent UI, after GPT-6.1 Sol arrived on September 29 at $2 / $10. See OpenAI rolls out GPT-6.
Dates to watch
| When | What |
|---|---|
| Oct 14 | Xiaomi routes MiMo-V2.5 requests to MiMo-V2.6-Pro; old model IDs stop working on Oct 21 |
| ”In October” | Kling 4.0 full model — no exact date, no API yet |
| End of October | Mistral Large 4 weights |
| Nov 17 | Google shuts down older Gemini TTS preview models (gemini-3.1-flash-tts-preview, gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts) |
| Jan 1, 2027 | Gemini 3.8 Flash TTS introductory pricing doubles |
Three trends from this stretch
- Rank and price have decoupled. The best-ranked image model and the best-value image model are different products, and the gap is about 6× per image. Choosing by leaderboard alone now costs real money.
- Voice is an agent feature. Realtime and speech-to-speech models, rather than plain text-to-speech, are where Alibaba and Amazon put their effort, and Microsoft shipped a low-latency Flash tier for agents. The pricing units are still inconsistent across vendors.
- “Announced” is not “available”. Kling 4.0 has only a Flash tier live, Mistral Large 4’s weights are weeks away, and Gemini 4 Argon access starts with a trusted-access program. We list these as previews and keep them out of rankings until they ship.
Caveats
- Arena scores for GPT Image 2.5 and Nano Banana 2.1 are preliminary and will move.
- Most voice-model claims are vendor-reported, and Alibaba’s price cuts are measured against its own earlier prices.
- Reports about the Gemini app’s free tier come from press coverage of Google’s help pages; check Google’s current pages before relying on them.
For model details, see each page linked above. For the broader LLM landscape, see the 2026 LLM guide.
Sources
- OpenAI — Introducing ChatGPT Images 2.5
- Google AI for Developers — Gemini Nano Banana 2.1
- Arena — Text-to-Image leaderboard
- Artificial Analysis — AA-Image-T2I v2.0 leaderboard
- Kling AI — Kling 4.0 vs 3.0
- Kling AI — API pricing
- Google AI for Developers — Deprecations
- Alibaba Cloud Model Studio — qwen-audio-3.1-realtime-plus