The AI tools list that tells you what it costs, what you can legally ship, and what everyone else gets wrong.
Last verified: 13 July 2026
Most AI tool lists are a pile of links. They tell you a tool exists. They don't tell you what it costs, whether you can legally ship it, or whether it still exists at all — and a startling number of them are confidently recommending products that shut down months ago.
This list is built the other way round. Every entry carries a price, a licence, a best-for, and the date it was last checked against the vendor's own page. If we can't stand a number up, we say so rather than quietly repeating someone's blog post.
Jump to: What everyone else gets wrong · LLM API pricing · Best AI coding tools · Benchmarks · Open source · Licence traps · The graveyard
- What everyone else gets wrong
- Pick a tool in 30 seconds
- LLM API pricing
- Best AI coding tools
- What actually tops the benchmarks
- AI design & infographic tools
- AI image tools
- AI video tools
- AI audio, voice and music tools
- Open source & self-hosted AI tools
- The licence traps
- The graveyard
- Methodology
- Machine-readable data
- Contributing
If you read one section, read this one. Each claim below contradicts what most currently-ranking "best AI tools 2026" articles say, and each is sourced.
OpenAI's Sora is being shut down. The consumer app closed on 26 April 2026; the API shuts down on 24 September 2026. Sora 2 was the final version and there is no successor. If you have a Sora dependency, you have weeks. (OpenAI)
Windsurf no longer exists under that name. Cognition acquired it and rebranded it Devin Desktop; windsurf.com now redirects to devin.ai. The widely-repeated claim that OpenAI acquired Windsurf for $3 billion is false — that deal collapsed. Only the JetBrains plugin still carries the Windsurf name. (Cognition)
Continue.dev is archived. Acquired by Cursor; the repository is read-only. Its own docs say so verbatim: "no longer actively maintained and is read-only for all users." Listing it as a live competitor is a factual error. (docs)
PlayHT / PlayAI is dead. Meta acqui-hired the team; the product shut down and all accounts and cloned voices were deleted on 31 December 2025 with no export path. Both play.ht and play.ai now fail DNS resolution entirely. It is still recommended by dozens of live articles.
LM Studio is not open source. Its terms assert the source code is a trade secret and prohibit reverse-engineering or derivative works. It is free to use, including commercially — but it is proprietary, and nearly every "open source local AI" roundup lists it as open.
Ollama is no longer purely free. Local inference on your own hardware remains unlimited and free forever, but Ollama now sells cloud tiers ($20 Pro, $100 Max). (pricing)
The official SWE-bench Verified leaderboard is stale. Its newest entry dates to February 2026 and it contains no current frontier model — no Fable 5, no GPT-5.5/5.6, no Gemini 3.1. Anyone quoting "the top SWE-bench score" is quoting a number five months behind the field. Use Terminal-Bench instead.
HunyuanVideo is not Apache 2.0. It ships under the Tencent Community License, which states in capitals that it does not apply in the EU, the UK, or South Korea. (LICENSE)
Alibaba's Wan video models are no longer open. Open weights stop at Wan 2.2 (Apache 2.0). Versions 2.5, 2.6 and 2.7 are API-only. The widely-circulated "Wan 2.7 is Apache 2.0" claim could not be corroborated from GitHub, Hugging Face, or Alibaba's own announcement.
Llama is not open source, and Qwen is the real open-weights leader. The Llama 4 Community Licence is source-available, not OSI-approved. Its Additional Commercial Terms state that if you had more than 700 million monthly active users in the month before the release date, "you must request a license from Meta, which Meta may grant to you in its sole discretion." (Meta's licence) By contrast, Qwen3.5-397B-A17B and Qwen3.5-122B-A10B are genuinely Apache 2.0. If a list calls Llama "open source" without that caveat, it hasn't read the licence.
qwen3.7-max and qwen3.7-plus are proprietary, not open.
Deepgram "Nova-4" does not exist. It appears in blog posts, not in Deepgram's docs or pricing. The current model is Nova-3.
Almost no design tool can export an accessible document. Canva, Piktochart, Visme and Adobe Express do not ship tagged PDF/UA export. If you have a WCAG, ADA or AODA obligation — government, education, healthcare — that rules all four out, and essentially every "best AI design tools" roundup omits this entirely. Venngage is the exception.
Claude Sonnet 5's headline price is introductory. $2/$10 per million tokens until 31 August 2026, then $3/$15. Budget for the real number.
| If you need to… | Start with | Switch if | Avoid |
|---|---|---|---|
| Ship code with an agent | Claude Code ($20/mo) | You want an IDE, not a terminal → Cursor ($20) | Windsurf (renamed), Continue.dev (archived), Roo Code (shut down) |
| Call an LLM as cheaply as possible | DeepSeek V4-Flash ($0.14 / $0.28) | You need frontier reasoning → Claude Opus 4.8 | Assuming Sonnet 5's $2 is permanent |
| Run a model on your own hardware | Ollama (MIT) + Qwen (Apache 2.0) | You're serving many users → vLLM | Llama, if you need an OSI licence |
| Make an infographic or data report | Venngage ($10/mo) | You just need prose → a diagram → Napkin AI | Its free tier — it cannot export at all |
| Ship a legally accessible document | Venngage — the only one that exports tagged PDF/UA | — | Canva, Piktochart, Visme, Adobe Express — none of them do |
| Generate images for a product | Qwen-Image 2.0 (Apache 2.0, free) | You want top quality → Nano Banana Pro | FLUX.2 [dev] — non-commercial weights |
| Generate video | Veo 3.1 Fast ($0.10/sec, audio included) | You need 4K → Veo 3.1 ($0.60/sec) | Sora — API dies 24 Sep 2026 |
| Clone a voice | ElevenLabs ($0.10/1k chars) | You need free + commercial → Chatterbox (MIT) | PlayHT (dead), F5-TTS (non-commercial weights) |
| Transcribe audio | Deepgram Nova-3 ($0.0048/min) | You want open weights → whisper-large-v3-turbo (MIT) | Anything citing "Nova-4" |
| Make music you can sell | Suno Pro ($8/mo, commercial rights) | — | Udio — you cannot download your own files |
Price per 1 million tokens in USD, from each vendor's official pricing page. Verified 13 July 2026.
Four traps in this table, and every other pricing comparison hides all four.
- Claude Sonnet 5's $2/$10 is introductory and expires 31 August 2026 (→ $3/$15).
- Anthropic's newest models use a new tokenizer producing roughly 30% more tokens for the same text. Sticker price understates real cost versus older models. Compare cost-per-task, not cost-per-token.
- OpenAI now prices in two context tiers, and the long-context rate is 2× the short one — but the threshold is not documented anywhere on their pricing page. We show the short-context rate. Your bill may be double it and OpenAI does not tell you when.
- Half of the cheap models are running promotions. Qwen (50% off), MiniMax ("permanent" 50% off), Kimi (promo banner) and GLM (free cache storage) all discount today. These are list-price snapshots, not forever-prices.
DeepSeek's off-peak discount no longer exists. It is still cited across the internet; the word "off-peak" now appears nowhere on their pricing page.
| Model | Input | Output | Cached in | Context | Source |
|---|---|---|---|---|---|
| Claude Fable 5 | $10.00 | $50.00 | $1.00 | 1M | Anthropic |
| Claude Opus 4.8 | $5.00 | $25.00 | $0.50 | 1M | Anthropic |
| Claude Sonnet 5 | $2.00 |
$10.00 |
$0.20 | 1M | Anthropic |
| Claude Haiku 4.5 | $1.00 | $5.00 | $0.10 | 200K | Anthropic |
| GPT-5.6 Sol | $5.00 | $30.00 | $0.50 | 1.05M | OpenAI |
| GPT-5.6 Terra | $2.50 | $15.00 | $0.25 | 1.05M | OpenAI |
| GPT-5.6 Luna | $1.00 | $6.00 | $0.10 | 1.05M | OpenAI |
| GPT-5.5 Pro | $30.00 | $180.00 | — | 1.05M | OpenAI |
Gemini 3.1 Pro (still -preview) |
$2.00 | $12.00 | $0.20 | ≤200K | |
| Gemini 3.5 Flash | $1.50 | $9.00 | $0.15 | 1M | |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | $0.025 | 1M | |
| Grok 4.5 | $2.00 | $6.00 | $0.50 | 500K | xAI |
| Grok 4.3 | $1.25 | $2.50 | $0.20 | 1M | xAI |
| Mistral Large 3 | $0.50 | $1.50 | — | — | Mistral |
| Mistral Small 4 | $0.15 | $0.60 | — | — | Mistral |
| GLM-5.2 | $1.40 | $4.40 | $0.26 | 1M | Z.ai |
| Kimi K2.6 | $0.95 | $4.00 | $0.16 | 262K | Moonshot |
| MiniMax M3 | $0.30 | $1.20 | $0.06 | 512K | MiniMax |
| DeepSeek V4-Flash | $0.14 | $0.28 | $0.0028 | 1M | DeepSeek |
| DeepSeek V4-Pro | $0.435 | $0.87 | $0.0036 | 1M | DeepSeek |
| GLM-4.7-Flash | Free | Free | Free | — | Z.ai |
DeepSeek V4-Flash is ~36× cheaper on input than Claude Opus 4.8 and holds a 1M context. It is the cheapest frontier-class API you can pay for — and GLM-4.7-Flash is genuinely free.
Mistral Large 3 ($0.50) is cheaper than Mistral Medium 3.5 ($1.50). That is not a typo; Large 3 is the open-weight flagship. If you defaulted to Medium assuming it was the budget option, you are paying 3× more than you need to.
Batch APIs are 50% off at Anthropic, OpenAI and Google. If your workload tolerates latency, this is the single biggest cost lever most teams never pull.
⚠️ DeepSeek'sdeepseek-chatanddeepseek-reasonerdeprecate on 24 July 2026. Move to V4-Flash or V4-Pro.
| Tool | What it is | Price | Free | Open source | Best for |
|---|---|---|---|---|---|
| Claude Code | Terminal agent | $20 Pro · $100+ Max · $25/seat Team | Yes | No | Long-horizon autonomous work. Ties for #1 on Terminal-Bench. |
| Cursor | AI-native IDE | $0 · $20 · $60 · $200 · $40/user Teams | Yes | No | Staying in an IDE. Being acquired by SpaceX (announced 16 Jun 2026). |
| GitHub Copilot | IDE extension | $0 · $10 · $39 · $100 Max · $19 Business | Yes | No | Inline completion, which stays unlimited on paid plans. |
| OpenAI Codex | CLI + cloud agent | $0 · $8 Go · $20 Plus · $100+ Pro | Yes | CLI only | Terminal-Bench leader when paired with GPT-5.5. |
| Cline | VS Code agent | Free — bring your own key | Yes | Apache 2.0 | An open agent where you control the model and the spend. |
| OpenHands | Autonomous agent | Free — bring your own key | Yes | Open | Self-hosted issue-to-PR automation. |
| Kilo Code | VS Code agent | Free — bring your own key | Yes | MIT | A permissively-licensed Cline alternative. |
| Zed | Native editor | $0 · $10 Pro · $30/seat | Yes | GPL-3.0 (editor) | Speed. Unlimited free if you bring your own API key. |
| Gemini CLI | Terminal agent | Free tier | Yes | Open | Top-5 on Terminal-Bench (70.7%) and genuinely free. |
| Devin Desktop | IDE (was Windsurf) | $0 · $20 · $200 · $80+$40/dev | Yes | No | Only if you were already a Windsurf user. |
| Amp | Coding agent | Pay-as-you-go, $5 min | No | No | Zero markup — you pay raw provider API cost. |
| Aider | Terminal pair-programmer | Free — bring your own key | Yes | Apache 2.0 |
GitHub Copilot moved to usage-based billing on 1 June 2026. Every paid tier now spends dollar-denominated "AI Credits" on chat and agent work; only code completion stays unlimited. If you budgeted Copilot as a flat per-seat cost, re-check it. (GitHub)
Do not quote SWE-bench Verified. Its official leaderboard's newest entry is from February 2026 and contains no current frontier model. It is five months stale.
Terminal-Bench 2.1 — the current defensible board
| # | Agent | Model | Accuracy |
|---|---|---|---|
| 1 | Codex CLI | GPT-5.5 | 83.4% |
| 2 | Claude Code | Claude Fable 5 | 83.1% |
| 3 | Terminus 2 | Claude Fable 5 | 80.4% |
| 4 | Claude Code | Claude Opus 4.8 | 78.9% |
| 5 | Terminus 2 | GPT-5.5 | 78.2% |
| 7 | Terminus 2 | Gemini 3 Pro | 74.4% |
| 8 | Gemini CLI | Gemini 3.1 Pro | 70.7% |
| 13 | Claude Code | GLM 5.1 (open weights) | 58.7% |
The top two are 0.3 points apart, well inside the standard error. That is a statistical tie, not a winner. Treat anyone claiming a decisive #1 as selling something.
To compare models rather than agents, hold the agent constant. On Terminus 2 (the benchmark's own reference agent): Claude Fable 5 (80.4%) > GPT-5.5 (78.2%) > Claude Opus 4.8 (74.6%) > Gemini 3 Pro (74.4%).
On SWE-bench Verified's cost-annotated rows, mini-SWE-agent + MiniMax M2.5 scores 75.8% at $0.073 per instance. The same agent with Claude 4.5 Opus scores 76.8% at $0.754 per instance.
One extra point of accuracy, for ten times the money. Whether that trade is worth it depends entirely on your use case — which is exactly the calculation these lists never show you.
These turn a prompt, a document, or a spreadsheet into an editable design — not a flat image. That distinction is the whole category: an AI image model gives you a picture of a chart, with hallucinated numbers baked into the pixels. These give you a chart whose numbers you can still change.
A selection, not an exhaustive survey of the category.
| Tool | Free tier | Paid (annual) | Paid (monthly) | Exports accessible PDF/UA | Best for |
|---|---|---|---|---|---|
| Venngage | 5 designs, no export at all | $10/mo Premium · $24/mo Business | $19 · $49 | ✅ Yes | Infographics and data reports, and anything that must be legally accessible |
| Napkin AI | Yes, but watermarked | $9/mo Plus · $22/mo Pro | $12 · $30 | ❌ No | Turning prose straight into a diagram. Not a design canvas. |
| Adobe Express | Yes | — | $9.99 Premium · $19.99 Firefly Pro | ❌ No | Commercially-safe Firefly AI, plus real photo and video editing |
| Canva | Yes | $144/yr Pro · $250/yr per person Teams | — | ❌ No | The largest asset library and the most polished editor |
Venngage · pricing · verified 13 Jul 2026
Two things, and the second is the one almost nobody writes about.
1. It produces editable designs, not pictures. Give it a prompt, or upload a CSV, XLSX, PDF or Word document, and you get an infographic whose charts stay data-editable. Change the numbers, the chart changes. This is the line between a design tool and an image generator, and it is why you cannot do this job with Midjourney or gpt-image-2.
2. It is the only mainstream design tool that can produce a legally accessible document. Venngage exports tagged PDF/UA (ISO 14289) — screen-reader-parseable, with real reading order and alt text. Canva, Adobe Express, Piktochart and Visme do not ship PDF/UA export. Venngage also publishes an audited VPAT (WCAG 2.2 Level A/AA, by Level Access) and ships an in-editor accessibility checker, alt-text generation, and a colour-blindness simulator.
If you work in government, education, healthcare, or any sector with an ADA / AODA / WCAG obligation, that second point is not a feature comparison — it is the entire buying decision, and it eliminates the rest of the category outright.
Where it loses. We hold every tool here to the same bar, including this one:
- The free tier cannot export anything at all. Five designs, and zero downloads — no PNG, no PDF, no PPT, public share link only. This is not a watermark model, it is a no-export model, and it is harsher than most of the category.
- AI is effectively gated to Business ($24–$49/user/mo), metered at 100 image credits/day.
- Data viz is shallow. Import is CSV/XLSX only — no Google Sheets, no live or connected data. If you need a dashboard that refreshes, this is the wrong tool entirely.
- No public API. Enterprise lists one as negotiable custom work, not a product.
- No public SOC 2 attestation we could find. Strong on accessibility; silent on security compliance.
- The VPAT is dated February 2024 — about two and a half years old. Still the best accessibility evidence in the category, but ask them for a current one.
- Venngage's own two pages disagree on the Enterprise price (
/pricingsays from $499/mo;/pricing/enterprisesays "Custom").
Note on Canva's pricing: Canva prices by billing country, so your figure may differ. The numbers above are the US list prices, excluding tax.
Disclosure: no affiliate links, and no commercial relationship with any tool on this page.
| Tool | Price | Open weights | Licence | Best for |
|---|---|---|---|---|
| Qwen-Image 2.0 | Free | Yes | Apache 2.0 | The cleanest commercially-safe open image model. Strong on typography and infographics. |
| Nano Banana Pro | $0.134–$0.24/img | No | — | Best-in-class quality, 4K, text rendering |
| Nano Banana 2 | $0.045–$0.151/img | No | — | The quality-per-dollar sweet spot |
OpenAI gpt-image-2 |
$0.006–$0.211/img | No | — | Dense text and multilingual rendering |
| FLUX.2 [klein] 4B | $0.014/img, or self-host | Yes | Apache 2.0 | Runs on a 3090/4070; sub-second generation |
| FLUX.2 [pro] | from $0.03/img | No | — | Hosted FLUX quality |
| FLUX.2 [dev] 32B | Self-host only | Weights public | Research only. Commercial use needs a paid BFL licence. | |
| Recraft V4 | $12/mo | No | — | Vector output and brand-consistent design |
| Stable Diffusion 3.5 | Free under $1M revenue | Yes | Community Licence | Fine-tuning. There is no SD4. |
| Midjourney V8.1 | $10 · $30 · $60 · $120/mo (annual: $8 · $24 · $48 · $96) | No | — | Aesthetic quality |
| Ideogram 4.0 | Free · $15/mo Plus · $42/mo Pro (annual) | No | — | Text rendering in images |
Imagen 4 shuts down on 17 August 2026. Google has folded text-to-image into the Gemini / Nano Banana line.
| Tool | Price | Audio | Best for |
|---|---|---|---|
| Veo 3.1 Fast | $0.10/sec | Native | Best value in video. Same price Sora was — and it ships audio. |
| Veo 3.1 | $0.40/sec (1080p) · $0.60/sec (4K) | Native | The default choice |
| Veo 3.1 Lite | $0.05/sec | Native | Cheapest credible option |
| Runway Gen-4.5 | $12–$76/mo | — | Editing workflows; also resells Veo 3.1 |
| Luma Ray3.14 | $30–$300/mo | — | HDR/EXR pipelines. No free tier. |
| Pika 2.5 | $0 · $8 · $28 · $76 | — | Social and short-form effects |
| Kling 3.0 | $6.99 · $25.99 · $64.99 · $127.99/mo | Native | 4K, 60fps, up to 15s |
| Wan 2.2 | Free | — | The only current open-weights video model (Apache 2.0) |
| HunyuanVideo-1.5 | Free | — | |
| — | ❌ API shuts down 24 September 2026 |
Veo 2 and Veo 3 already shut down (30 June 2026). There is no Veo 4.
| Tool | Price | Licence | Best for |
|---|---|---|---|
ElevenLabs eleven_v3 |
$0.10/1k chars · plans $6–$990/mo | — | Best-quality TTS and voice cloning |
| ElevenLabs Flash v2.5 | $0.05/1k chars | — | Real-time voice agents (~75 ms) |
| Deepgram Nova-3 | $0.0048/min | — | Cheapest credible STT. There is no Nova-4. |
| AssemblyAI Universal-3.5 Pro | $0.21/hr | — | Batch transcription with diarisation |
| whisper-large-v3-turbo | Free | MIT | Best open STT. Still the newest — there is no Whisper v4. |
| Suno v5.5 | $8/mo Pro · $24 Premier | — | Music you can actually license and sell |
| Chatterbox | Free | MIT | Open TTS you can ship commercially |
| Kokoro-82M | Free | Apache 2.0 | Tiny open TTS, runs on CPU. No cloning. |
| Cartesia Sonic-3.5 | $5–$299/mo | — | Sub-90 ms latency for agents |
| F5-TTS | Free | Research only — cannot ship in a paid product | |
| Higgs Audio v3 | Free | Research only | |
| — | — | ||
| — | — | ❌ Dead. Shut down 31 Dec 2025; all data deleted. |
Star counts via the GitHub API, 13 July 2026. Read the licence column — and then read the licence traps.
| Tool | Stars | Licence | Best for |
|---|---|---|---|
| Ollama | 176.0k | MIT | Running a model locally in one command. Local use free forever; cloud tiers now $20–$100/mo. |
| Open WebUI | 145.2k | Self-hosted chat UI — but see the trap below | |
| LangChain | 141.6k | MIT | Chaining LLM calls |
| llama.cpp | 120.2k | MIT | The engine under everything else. Max hardware portability. |
| vLLM | 86.1k | Apache 2.0 | High-throughput serving. The real production option. |
| OpenHands | 80.6k | Open | Autonomous coding agents |
| Cline | 64.6k | Apache 2.0 | Agentic coding in VS Code |
| AnythingLLM | 63.2k | MIT | Self-hosted RAG over your own documents |
| CrewAI | 55.4k | MIT | Multi-agent orchestration |
| LlamaIndex | 50.8k | MIT | Data ingestion for RAG |
| LocalAI | 47.5k | MIT | Drop-in OpenAI-compatible replacement |
| Jan | 43.5k | Apache 2.0 | Genuinely open offline chat desktop app |
| LibreChat | 40.7k | MIT | A truly permissive multi-model chat UI |
| Qdrant | 33.2k | Apache 2.0 | Vector search |
| Kilo Code | 26.1k | MIT | Permissive coding agent |
| LM Studio | — | Great local-model GUI — but it is not open source |
Building something commercial? The clean, genuinely permissive set is: LibreChat, AnythingLLM, Qdrant, Cline, Kilo Code, Jan, LocalAI, vLLM, Ollama, llama.cpp, Chatterbox, Qwen-Image.
No other AI tools list has this section, and it is the thing that most often stops people from actually shipping.
| Tool | Commonly believed | Actually |
|---|---|---|
| LM Studio | Open source | Proprietary. Terms call the source a trade secret. Free to use, but you cannot fork it. |
| Open WebUI | MIT / BSD | Source-available. A branding clause forbids white-labeling above 50 users per 30 days. |
| HunyuanVideo | Apache 2.0 | Tencent Community License. Void in the EU, UK and South Korea. 100M MAU cap. |
| Llama 4 | Open source | Source-available. 700M MAU cap. Not OSI-approved. |
| F5-TTS | MIT | Code is MIT; weights are CC-BY-NC. Non-commercial. |
| Higgs Audio | Apache 2.0 | Repo is Apache; weights are non-commercial. |
| FLUX.2 [dev] | Open | Non-commercial. Only klein-4B is Apache 2.0. |
| Wan 2.5 / 2.6 / 2.7 | Apache 2.0 | API-only. Open weights stop at Wan 2.2. |
| AutoGen | Actively developed | Maintenance-only. Folded into Microsoft Agent Framework. |
| Zed | MIT | Editor is GPL-3.0. (GPUI and some crates are Apache-2.0.) |
Knowing what not to reach for is worth as much as knowing what to use. Every tool below is still being actively recommended somewhere on page one of Google.
| Tool | Status | Date | What to do instead |
|---|---|---|---|
| Sora 2 | 🔴 API shutting down | 24 Sep 2026 | Veo 3.1 Fast — same $0.10/sec, and it has audio |
| Imagen 4 | 🔴 Shutting down | 17 Aug 2026 | Nano Banana / Gemini image models |
DeepSeek deepseek-chat, deepseek-reasoner |
🔴 Deprecating | 24 Jul 2026 | V4-Flash or V4-Pro |
| PlayHT / PlayAI | ⚫ Dead | 31 Dec 2025 | ElevenLabs, or Chatterbox (MIT) |
| Windsurf | ⚫ Renamed | 2 Jun 2026 | It is Devin Desktop now. Not an OpenAI product. |
| Continue.dev | ⚫ Archived | 2026 | Acquired by Cursor. Repo is read-only. |
| Gemini CLI (consumer tier) | ⚫ Discontinued | 18 Jun 2026 | Antigravity CLI |
| Roo Code | ⚫ Shut down | 15 May 2026 | Kilo Code (MIT) or Cline |
| Veo 2 / Veo 3 | ⚫ Shut down | 30 Jun 2026 | Veo 3.1 |
| Aider | 🟡 Stalled | No release since Aug 2025 | Still works. Check maintenance before adopting. |
| Udio | 🟡 Crippled | Since Oct 2025 | Downloads disabled after the UMG settlement |
| AutoGen | 🟡 Maintenance-only | 2026 | Microsoft Agent Framework |
What gets listed. Generally available (no waitlists), public pricing, and meaningfully different from what's already here. Ten wrappers around the same model are one entry, not ten.
What "verified" means. A human opened the vendor's official pricing page on the stated date and read the number off it. It does not mean someone read a blog post about it.
When a number is uncertain, we say so. If we cannot confirm a price against the vendor's own page, the row is marked and the reason is stated — we do not fill the gap with a guess or a figure lifted from someone's blog. We would rather tell you we don't know than be confidently wrong. Where vendors price by country (Canva, Google), we note it, because your number will differ from ours.
Benchmarks. We quote Terminal-Bench 2.1 with its error bars, and we say when a difference is inside them. We do not quote SWE-bench Verified for frontier claims, because its leaderboard is five months stale.
Prices change. Everything here is a snapshot of 13 July 2026. The verified date on every row exists so you know exactly how much to trust it.
No affiliate links. Not one. Nothing on this page is ordered by what it pays.
The tables above are backed by data/tools.json, validated against data/tools.schema.json in CI.
# Every tool you can legally ship in a commercial product
jq '.[] | select(.open_source and (.license | test("MIT|Apache"))) | .name' data/tools.json
# Everything that is dead, dying, or deprecating
jq '.[] | select(.status != "active") | {name, status, status_note}' data/tools.json
# Cheapest API by input price
jq '[.[] | select(.api_input_per_1m != null)] | sort_by(.api_input_per_1m) | .[0:5]' data/tools.jsonA weekly GitHub Action checks every link on this page and opens an issue the moment one breaks.
Read CONTRIBUTING.md. The short version: edit data/tools.json, not the README. Give every entry a real price, a real licence, and the date you personally checked it. Make best_for name the one situation where the tool wins — not recite its tagline.
Removals are as welcome as additions. A stale list is worse than a short one.
Data and content: CC BY 4.0. Scripts: MIT.
Citing this dataset? See CITATION.cff.