The best AI model for every job in 2026.
The lowest cost, best value and premium pick for chat, voice, transcription, image and video. Worked out from current prices on each provider's own page, by the same engine that runs our Cost Comparator.
One pick per job, at three price points.
| Job | Lowest cost | Best value | Premium |
|---|---|---|---|
| Text and chat | GPT OSS 20B (Groq)$0.13 per million tokens, blended | Claude Haiku 5.5$0.20 per million tokens, blended | Claude Sonnet 5.5$4.00 per million tokens, blended |
| Text to speech | Azure Neural TTS$0.015 a minute | Azure Neural TTS$0.015 a minute | ElevenLabs Multilingual v2$0.080 a minute |
| Speech to text | AssemblyAI Universal-2$0.0025 a minute | AssemblyAI Universal-2$0.0025 a minute | AssemblyAI Universal-2$0.0025 a minute |
| Image generation | GPT Image 2.5 (medium)$0.013 an image | GPT Image 2.5 (medium)$0.013 an image | FLUX.2 [pro]$0.030 an image |
| Video generation | Runway Gen-4 Turbo$0.050 a second | Runway Gen-4 Turbo$0.050 a second | Kling 3.0 (720p)$0.084 a second |
The best AI model for text and chat.
Chat, agents, extraction and anything else that reads and writes text. At 150,000+ calls/month, every dollar per million tokens matters. Route the routine 80% through a cheap fast model and escalate the hard 20% to premium. Caching alone can cut this bill by 30 to 50%. Monthly figures below assume 1,000 users a day making five requests each, 1,500 tokens in and 500 out.
GPT OSS 20B (Groq)
- At the reference workload
- $39 a month
- Our tier
- Basic
Claude Haiku 5.5
- At the reference workload
- $60 a month
- Our tier
- Standard
Prompts over 100,000 tokens cost $0.50 input and $2.50 output. Counts about 30 percent more tokens than Haiku 4.5 for the same text
Claude Sonnet 5.5
- At the reference workload
- $1,200 a month
- Our tier
- Premium
Cache reads $0.10 (cut from $0.20 on 7 October 2026). Uses Sonnet 5 tokenizer: about 30% more tokens than Sonnet 4.6, 4.5 or Haiku 4.5 for the same text, so compare on token counts, not rates
| Model | Tier | Price | A month here |
|---|---|---|---|
| GPT OSS 20B (Groq)Groq | Basic | $0.13 per million tokens, blended | $39 |
| Claude Haiku 5.5Anthropic | Standard | $0.20 per million tokens, blended | $60 |
| GPT-6 LunaOpenAI | Standard | $0.20 per million tokens, blended | $60 |
| DeepSeek V4.1 FlashDeepSeek | Standard | $0.26 per million tokens, blended | $79 |
| GPT OSS 120B (Groq)Groq | Standard | $0.26 per million tokens, blended | $79 |
| GPT-5.6 LunaOpenAI | Standard | $0.45 per million tokens, blended | $135 |
| Mistral LargeMistral | Premium | $0.75 per million tokens, blended | $225 |
| Gemini 3.5 Flash LiteGoogle | Basic | $0.85 per million tokens, blended | $255 |
| DeepSeek V4 ProDeepSeek | Premium | $0.99 per million tokens, blended | $297 |
| Sonar (web aware)Perplexity | Standard | $1.00 per million tokens, blended | $300 |
| Llama 3.3 70B (Together)Together | Standard | $1.04 per million tokens, blended | $312 |
| Gemini 3.7 FlashGoogle | Standard | $1.50 per million tokens, blended | $450 |
- Cache responses aggressively. Same prompt in = same output.
- Use a small model as a router for a big model. Cheap triage before expensive reasoning.
- Stream tokens. Perceived speed matters more than actual speed.
- Set per user rate limits so one bug does not become a bankruptcy event.
The best text to speech API.
Voice agents, narration and read aloud features. Voice is billed by the minute of audio you generate, so the gap between a standard and a premium voice compounds with every listener. Monthly figures below assume 500 minutes of generated speech a day.
Azure Neural TTS
- At the reference workload
- $225 a month
- Our tier
- Standard
$15 per million characters, East US pay as you go
ElevenLabs Multilingual v2
- At the reference workload
- $1,200 a month
- Our tier
- Premium
$0.08 per 1,000 characters, about one minute of speech
| Model | Tier | Price | A month here |
|---|---|---|---|
| Azure Neural TTSAzure | Standard | $0.015 a minute | $225 |
| Deepgram AuraDeepgram | Standard | $0.015 a minute | $225 |
| Google Cloud TTS (Neural2)Google | Standard | $0.016 a minute | $240 |
| ElevenLabs Flash / TurboElevenLabs | Standard | $0.040 a minute | $600 |
| ElevenLabs Multilingual v2ElevenLabs | Premium | $0.080 a minute | $1,200 |
| Google Cloud TTS (Studio)Google | Premium | $0.16 a minute | $2,400 |
- Cache generated audio by (voice + text hash). Same text = same file.
- For short utterances (< 30 chars), the extra latency of premium voices rarely justifies the cost.
- Use SSML for control over pacing, emphasis, pauses. Cheap voices sound better with good SSML.
- Chunk long text and generate in parallel. Users hear the first chunk before the rest finishes.
The best speech to text API.
Transcribing calls, meetings and voice input. Transcription is billed by the minute of audio you send, and the current models are cheap enough that accuracy on your own audio should decide it. Monthly figures below assume 500 minutes of audio transcribed a day.
AssemblyAI Universal-2
- At the reference workload
- $38 a month
- Our tier
- Premium
$0.15 an hour. Universal-3.5 Pro is $0.21 an hour
| Model | Tier | Price | A month here |
|---|---|---|---|
| AssemblyAI Universal-2AssemblyAI | Premium | $0.0025 a minute | $38 |
| Deepgram Nova 3Deepgram | Premium | $0.0043 a minute | $65 |
| Google Cloud STTGoogle | Standard | $0.016 a minute | $240 |
| Azure Speech to TextAzure | Standard | $0.017 a minute | $255 |
- Stream input. Users see partial transcripts as they speak, and far fewer drop off.
- Use language detection only if you need it. Explicit language saves cost and improves accuracy.
- For call recording, chunk the audio into five minute segments. Providers charge per minute of audio input.
- Store the original audio + transcript for training your own retrieval later.
The best image generation API.
Product shots, illustrations, avatars and marketing images. Images are billed per image, and resolution and quality settings move the price as much as the choice of model does. Monthly figures below assume 200 images a day.
GPT Image 2.5 (medium)
- At the reference workload
- $79 a month
- Our tier
- Standard
Priced per token: about 439 output tokens for a 1024x1024 medium image at $30 per million, plus prompt tokens. High quality is about $0.053
FLUX.2 [pro]
- At the reference workload
- $180 a month
- Our tier
- Premium
Per 1 megapixel image
| Model | Tier | Price | A month here |
|---|---|---|---|
| GPT Image 2.5 (medium)OpenAI | Standard | $0.013 an image | $79 |
| FLUX.2 [klein] 4BBFL | Basic | $0.014 an image | $84 |
| FLUX.2 [pro]BFL | Premium | $0.030 an image | $180 |
| Recraft v3Recraft | Premium | $0.040 an image | $240 |
| Stable Diffusion 3.5Stability | Premium | $0.065 an image | $390 |
| Ideogram v2Ideogram | Premium | $0.080 an image | $480 |
| Ideogram v3Ideogram | Premium | $0.090 an image | $540 |
- Cache the (prompt + seed) combination. Same input = same image, no need to regenerate.
- Use a cheap model (FLUX.2 [klein]) for drafts. A stronger one (FLUX.2 [pro], Recraft, Ideogram) for final assets.
- Optimize output size at the source: request 1024 not 2048 if you'll downsize.
- Generate variations ahead of time for common queries and store in a CDN.
The best video generation API.
Short clips, ads and social video. Video is billed per second of output and costs far more than any other category, so cap generations before anything else. Monthly figures below assume 60 seconds of video a day.
Runway Gen-4 Turbo
- At the reference workload
- $90 a month
- Our tier
- Standard
5 credits a second at $0.01 a credit
Kling 3.0 (720p)
- At the reference workload
- $151 a month
- Our tier
- Premium
0.6 units a second at $0.14 a unit, without native audio
| Model | Tier | Price | A month here |
|---|---|---|---|
| Runway Gen-4 TurboRunway | Standard | $0.050 a second | $90 |
| Luma Ray3.2 (720p)Luma | Standard | $0.060 a second | $108 |
| Kling 3.0 (720p)Kling | Premium | $0.084 a second | $151 |
| Runway Gen-4.5Runway | Premium | $0.12 a second | $216 |
- Video is the most expensive category by orders of magnitude. Cache aggressively.
- Use shorter durations (2 to 4s) for social loops. Longer is exponentially more expensive.
- Preview with cheap models before committing to premium generation.
- Consider generating stills + Ken Burns effect for pseudo video. 90% of the wow, 5% of the cost.
How the picks are made.
Lowest cost is the cheapest current model. Best value is the cheapest current model outside our basic tier. Premium is the cheapest current model in our premium tier. Tiers are our grouping by price band and the job each provider positions a model for, not a benchmark score.
Text picks come from the routing engine: a cheap model for volume, a workhorse for routine work and a stronger model for hard cases, chosen by monthly call volume. At the reference volume of 150,000 requests a month, that means a cheap model for volume, a workhorse for routine work and a stronger model for the hard cases.
Only current models are picked. Legacy, retired and delisted models are left out unless nothing else is selected. Prices are list prices from each provider’s own pricing page, with no caching or batch discounts, reconciled by hand every month and checked weekly by an automated drift check. The roster behind this page is dated 23 Oct 2026.
The same engine runs the recommendations in the Cost Comparator, applied there to the models you select at your own volume. The two cannot disagree, because they are one piece of code.
Free to reuse, with credit.
The picks and every current price on this page are published as JSON at https://pulse.cafiyn.com/free/best-ai-models.json, rebuilt with the site. Use it in your own tools, docs or agents under CC BY 4.0: credit Cafiyn Pulse with a link to this page.
Common questions.
What is the best text and chat model in 2026?
Lowest cost: GPT OSS 20B (Groq) from Groq at $0.13 per million tokens, blended, about $39 a month at 1,000 users a day making five requests each, 1,500 tokens in and 500 out. Best value: Claude Haiku 5.5 from Anthropic at $0.20 per million tokens, blended, about $60 a month at 1,000 users a day making five requests each, 1,500 tokens in and 500 out. Premium: Claude Sonnet 5.5 from Anthropic at $4.00 per million tokens, blended, about $1,200 a month at 1,000 users a day making five requests each, 1,500 tokens in and 500 out. These are worked out from current list prices on each provider's own page, dated 23 Oct 2026.
What is the best text to speech model in 2026?
Lowest cost and best value: Azure Neural TTS from Azure at $0.015 a minute, about $225 a month at 500 minutes of generated speech a day. Premium: ElevenLabs Multilingual v2 from ElevenLabs at $0.080 a minute, about $1,200 a month at 500 minutes of generated speech a day. These are worked out from current list prices on each provider's own page, dated 23 Oct 2026.
What is the best speech to text model in 2026?
Lowest cost and best value and premium: AssemblyAI Universal-2 from AssemblyAI at $0.0025 a minute, about $38 a month at 500 minutes of audio transcribed a day. These are worked out from current list prices on each provider's own page, dated 23 Oct 2026.
What is the best image generation model in 2026?
Lowest cost and best value: GPT Image 2.5 (medium) from OpenAI at $0.013 an image, about $79 a month at 200 images a day. Premium: FLUX.2 [pro] from BFL at $0.030 an image, about $180 a month at 200 images a day. These are worked out from current list prices on each provider's own page, dated 23 Oct 2026.
What is the best video generation model in 2026?
Lowest cost and best value: Runway Gen-4 Turbo from Runway at $0.050 a second, about $90 a month at 60 seconds of video a day. Premium: Kling 3.0 (720p) from Kling at $0.084 a second, about $151 a month at 60 seconds of video a day. These are worked out from current list prices on each provider's own page, dated 23 Oct 2026.
How are the picks chosen?
Lowest cost is the cheapest current model. Best value is the cheapest current model outside our basic tier. Premium is the cheapest current model in our premium tier. Tiers are our grouping by price band and the job each provider positions a model for, not a benchmark score. Text picks come from the routing engine: a cheap model for volume, a workhorse for routine work and a stronger model for hard cases, chosen by monthly call volume. Only current models are picked. Legacy, retired and delisted models are left out unless nothing else is selected.
Is the cheapest model good enough?
Often, for high volume work such as classification, routing, short answers and drafts. The way to know is to run your own inputs through two or three candidates and compare the results. The Cost Comparator prices each candidate at your own volume.
How often is this updated?
The picks are computed from the roster behind the Cost Comparator every time the site is built. That roster is checked weekly by an automated drift check and reconciled by hand every month; the current version is dated 23 Oct 2026. Every price change is dated and sourced in the price change log.
Can I use this data in my own tool?
Yes. The picks and every current price are published as JSON at https://pulse.cafiyn.com/free/best-ai-models.json under CC BY 4.0. Credit Cafiyn Pulse with a link to this page.
Pick on your own numbers.
Set your users, requests, audio minutes, images and video seconds, select the models you are weighing, and get a pick for every part of your product.
Price my own usage