July 2026 Edition, Verified
The Complete Guide to Free LLM APIs
and Low-Cost Inference in 2026
Token costs are climbing. Your Hermes Agent, OpenClaw, or Open Code project should not stall because of a billing alert. Here is every provider offering free API access right now, what the real limits are, and the cheapest unlimited-style subscriptions when you need to go further.
Why this matters: A multi-step agent workflow can run $40 to $200 a month at list price. Stacking free tiers intelligently covers 80 to 90 percent of that for $0. The remaining calls, where you actually need frontier-level reasoning, is where you spend. This guide shows you both sides.
Quick Glossary: RPM Means Requests Per Minute, RPD Means Requests Per Day
Terms you will see throughout this guide
RPM = Requests Per Minute
The number of API calls you can send in any 60-second window before the provider throttles or rejects the next one with a 429 error.
The number of API calls you can send in any 60-second window before the provider throttles or rejects the next one with a 429 error.
RPD = Requests Per Day
The total number of API calls allowed in a rolling or calendar day, regardless of how you space them out.
The total number of API calls allowed in a rolling or calendar day, regardless of how you space them out.
TPM
Tokens Per Minute. A separate ceiling some providers use instead of, or alongside, RPM, since one request can contain very different amounts of text.
Tokens Per Minute. A separate ceiling some providers use instead of, or alongside, RPM, since one request can contain very different amounts of text.
429 Error
The HTTP status code returned when you exceed a rate limit. It means slow down, not you are banned.
The HTTP status code returned when you exceed a rate limit. It means slow down, not you are banned.
Why the units matter: A provider advertising "1,000/day" and one advertising "1,000/month" are not comparable. Throughout this guide, every number is labeled RPD (per day) or RPM (per minute) so you can compare providers on equal footing. Where a provider sets different caps per model, we list the specific caps for each top model rather than a single blended number.
Lean Team, Building Fast
Jump to the Recommended Stack. Google AI Studio + Cerebras + OpenRouter covers most MVPs at $0.
Running Hermes Agents
See the NVIDIA NIM card. It now hosts GLM-5.2 and Nemotron 3 Ultra free, purpose-built for agent frameworks.
Scaling Past Free Tiers
Featherless and the GLM Coding Plan give flat-rate unlimited usage starting at $10 to $18/month.
Coding in Cursor or VS Code
See what Cursor, Codex, and Antigravity actually give you free before you touch a raw API.
Quick Reference: All Free API Tiers
No credit card, no trial period limits, verified July 4, 2026
| Provider | Top Free Models (Verified) | Daily / Monthly Limit | Rate Limit | Card Required | Specialty Modalities | How Long Free |
|---|---|---|---|---|---|---|
| Google AI Studio | Gemini 3.5 Flash, Gemini 3.1 Pro (preview), Gemini 3 Flash-Lite | ~1,500 RPD on Flash-Lite; lower on 3.5 Flash | 10-15 RPM | No | Vision, Audio, Video, Deep Research (free) | Ongoing; quotas cut in late 2025 |
| NVIDIA NIM | GLM-5.2, Nemotron 3 Ultra, Llama 4 Scout | No credit or token cap; access is gated entirely by rate limit | 40 RPM hard cap, shared across all models, non-negotiable on free tier | No | Vision, embeddings, safety/guard models | Permanently free; NVIDIA removed its earlier 5,000-credit ceiling in 2026 |
| Groq | Llama 4 Scout, Llama 3.3 70B, Kimi K2 | 1,000-14,400 req/day | 30 RPM, 6-30K TPM | No | Whisper Large v3 & Turbo STT (free) | Free Forever tier |
| Cerebras | Llama 3.3 70B, Llama 4 Scout, Qwen3-32B | Model-dependent daily budget | 30 RPM / 60K TPM | No | Fastest raw inference speed (2,000+ tok/sec) | Ongoing free tier |
| OpenRouter | DeepSeek R1, Qwen3 Coder 480B, GLM-5.2 | 50 requests/day (no balance); 1,000 requests/day (with $10+ lifetime credit purchase) | 20 RPM | No | Vision and tool calling on select free models | Ongoing; 25+ free models rotate, platform-wide cap applies to all of them equally |
| Mistral La Plateforme | Mistral Small 3.2, Devstral Small, Codestral | 1 req/sec (Experiment tier) | ~1 RPS | Phone verify | Vision (Pixtral), Code (Codestral) | Ongoing Experiment tier |
| GitHub Models | GPT-5-mini, Llama 4 Scout, DeepSeek-R1 | 10-150 req/day (model class dependent) | 10-15 RPM | No | Vision on GPT-4.1 / o4-mini class models | Ongoing for GitHub account holders |
| SambaNova Cloud | Llama 3.3 70B, DeepSeek-V3.1, gpt-oss-120b | Model-specific RPM caps | 10-30 RPM | No | Llama Vision (rate limited) | Ongoing free tier |
| Hugging Face Inference | Llama 3.2 11B Vision, Qwen2.5, Flux.1 Dev | Credit-based monthly quota | Cold-start dependent | No | Image gen, audio, STT | Ongoing (shared infra) |
| Cohere | Command A, Command R7B, Rerank 3.5 | 1,000 API calls/month | 20 RPM | No | Embed v4, Rerank, STT (5 RPM) | Ongoing, non-commercial only |
| DeepSeek | DeepSeek-V4-Flash, DeepSeek-R1 | Signup credit varies by promo | Low RPM without paid tier | After credit | 1M context, reasoning mode | One-time credit, then pay-as-you-go |
| Cloudflare Workers AI | Llama 4 Scout, Mistral 7B, Gemma 7B | 10,000 neurons/day | Standard Workers limits | No | Whisper STT, text-to-image, embeddings | Ongoing free plan |
Important: Free tier limits and model names change constantly. Verify current model IDs and quotas on the provider's own docs page before wiring anything into production. RPM and RPD reflect per-model caps published by each provider, not one blanket number.
Provider Deep Dives
Ranked using OpenRouter usage and benchmark data where applicable
NVIDIA NIM
build.nvidia.com
NVIDIA hosts genuinely frontier-class open weight models on its own silicon, including Zhipu AI's GLM-5.2, to prove out NIM as the default runtime for open models. NVIDIA removed its earlier consumable-credit system in 2026, so the free tier is now permanent and gated only by rate limit, not by a shrinking credit balance.
Free Usage Cap
None. No credits, no token ceiling, no monthly quota
Rate Limit
40 RPM hard cap, shared across every model in the catalog
GLM-5.2 Context Window
Up to 1,000,000 tokens (model card spec)
Duration
Permanently free; the 40 RPM ceiling is not raisable on the free tier
z-ai/glm-5.2nvidia/nemotron-3-ultrameta/llama-4-scout
GLM-5.2 and Nemotron 3 Ultra rank highest on NVIDIA's own catalog usage; DeepSeek-R1 remains available but is not consistently a top-3 model on the NIM catalog as of July 2026.
Vision, EmbeddingsData may be used for training
Access: build.nvidia.com/z-ai/glm-5.2. Sign up for a free permanent API key, no credit card and no credit balance to track. The 40 RPM cap applies across all models in the catalog and cannot be raised on the free tier, so it is the only real constraint. An agent's iterative tool calls can still burn through 40 RPM fast, leading to lag or 429 errors on complex tasks.
Do not use for confidential or client data. NVIDIA's free-tier terms permit using inputs and outputs to improve their own models.
Do not use for confidential or client data. NVIDIA's free-tier terms permit using inputs and outputs to improve their own models.
Google AI Studio
ai.google.dev
Google gives this away to convert developers onto Vertex AI billing. Gemini 3.5 Flash, released May 19, 2026 at I/O, is now the flagship free-tier model and is genuinely frontier-class for agents and coding, per Google's own benchmarks against Gemini 3.1 Pro.
Gemini 3.5 Flash
Lower RPM, priority quota
Gemini 3 Flash-Lite
Up to 1,500 requests per day (RPD)
Context Window
1,048,576 tokens on Flash-class models
Multimodal
Text, Vision, Audio, Video, PDF
gemini-3.5-flashgemini-3.1-pro (preview)gemini-3-flash-lite
Gemini 3.5 Flash reportedly beats Gemini 3.1 Pro on most agentic and coding benchmarks while running roughly 4x faster. Free-tier RPM/RPD is model-specific: lighter models like Flash-Lite get the highest daily caps, while 3.5 Flash and the 3.1 Pro preview carry tighter per-minute limits.
VisionAudio InputVideo Input
Access: ai.google.dev → Get API Key. Google account only, no card.
Watch out: Google reduced free quotas in late 2025 and may again; confirm current per-model RPM/RPD on the AI Studio rate limits page before relying on a specific number.
Watch out: Google reduced free quotas in late 2025 and may again; confirm current per-model RPM/RPD on the AI Studio rate limits page before relying on a specific number.
Groq
console.groq.com
Groq built custom LPU chips and needs developers to feel the speed advantage firsthand. Once you experience 300+ tokens/second, slower frontier APIs feel sluggish. Their hardware business is the real product.
Best RPD
14,400 (Llama 3.1 8B)
70B Model RPD
1,000/day
Speed
300+ tokens/sec
Rate Limit
30 RPM
meta-llama/llama-4-scoutllama-3.3-70b-versatilemoonshotai/kimi-k2-instruct
Whisper Large v3 STT (Free)Ultra-Low Latency
Access: console.groq.com → Sign up → API Keys.
Best for: Voice agents and real-time chat. Kimi K2 remains a strong free option for agentic tool use.
Best for: Voice agents and real-time chat. Kimi K2 remains a strong free option for agentic tool use.
Cerebras
cloud.cerebras.ai
Cerebras uses wafer-scale chips and needs workloads that showcase raw inference speed, not necessarily a higher daily request cap than everyone else. Its edge is throughput per request: it generates tokens faster than any other free provider on this list.
Speed
2,000+ tokens/sec
Rate Limit
30 RPM / 60K TPM
Context
Long context capable
Card
Not required
llama-4-scoutllama-3.3-70bqwen3-32b
Correction: Cerebras' advantage is raw speed, not a bigger daily allowance than Google's request-based quota. Compare by what actually constrains your workload.
Fastest Free InferenceLong Context RAG
Access: cloud.cerebras.ai → Sign up. Email only.
Best for: Batch document processing where wall-clock speed, not daily volume, is the bottleneck.
Best for: Batch document processing where wall-clock speed, not daily volume, is the bottleneck.
OpenRouter
openrouter.ai
OpenRouter monetizes by routing paid calls across providers and taking a small margin. Free models are subsidized by labs who want developer adoption. One API key for 25+ rotating free models is the fastest way to test many models side by side.
Free Requests, No Balance
50 requests per day (RPD)
Free Requests, $10+ Balance
1,000 requests per day (RPD)
Rate Limit
20 requests per minute (RPM)
API Style
OpenAI-compatible
deepseek/deepseek-r1:freeqwen/qwen3-coder-480b:freeopenai/gpt-oss-120b:free
Top 3 based on OpenRouter's most-used free models for coding and agent tasks as of July 2026. Failed requests still count against your daily quota.
Vision (select models)Tool Calling (select models)
Access: openrouter.ai → Sign up → API Keys. Use any model ID ending in
:free, or see the Free Router section below. The 50 vs. 1,000 RPD split applies to your total free-model usage across ALL free models combined, not per model.Mistral La Plateforme
console.mistral.ai
Mistral's Experiment tier is top-of-funnel for a European lab competing with OpenAI and Anthropic. Startups can also apply for grants up to $30,000 in credits through Mistralship.
Rate Limit
~1 req/sec
All API Models
Yes, incl. Codestral
EU Data Residency
Yes
Startup Grant
Up to $30K via Mistralship
mistral-small-3.2devstral-smallcodestral
Vision (Pixtral)EU GDPR Compliant
Access: console.mistral.ai → Sign up → Experiment tier. Phone SMS verification required.
GitHub Models
github.com/marketplace/models
Microsoft owns GitHub and Azure OpenAI. GitHub Models is the discovery layer that funnels production usage toward Azure.
Total Models
45+
High-Tier RPD
10-50/day
Low-Tier RPD
up to 150/day
Per-Request Limit
Model dependent
openai/gpt-5-minimeta/llama-4-scoutdeepseek/deepseek-r1
Vision (GPT class)OpenAI-Compatible
Access: github.com/marketplace/models → Generate a token with Models:read scope. Base URL:
https://models.inference.ai.azure.comSambaNova Cloud
cloud.sambanova.ai
SambaNova built RDU chips as an NVIDIA alternative. Rate limits are per-model rather than a single flat cap, so smaller models get far more headroom than 405B-class ones.
Small Models
Up to 30 RPM
70B Models
~20 RPM
405B Models
~10 RPM
Speed
RDU-accelerated
Meta-Llama-3.3-70B-InstructDeepSeek-V3.1gpt-oss-120b
Llama Vision (rate limited)OpenAI-Compatible
Access: cloud.sambanova.ai → Sign up → API Key (starts with
sn-).Hugging Face Inference
huggingface.co
HuggingFace is the GitHub of AI models. Free inference keeps developers on-platform, driving traffic toward paid Inference Endpoints and PRO plans.
Access Type
Credit-based serverless
Model Variety
Thousands
API Format
HF Inference (not OAI)
Cold Starts
Common on rare models
Llama-3.2-11B-Vision-InstructQwen2.5-72B-InstructFLUX.1-dev
Image Gen (Flux.1 Dev)Audio / STT
Access: huggingface.co → Settings → Access Tokens → Create read token. Not OpenAI-compatible.
Cohere
dashboard.cohere.com
Cohere is enterprise-focused and uses the free trial key to let developers validate RAG pipeline quality before committing. Rerank and Embed remain best-in-class for retrieval at zero cost.
Monthly Calls
1,000 calls/month
Rate Limit
20 RPM
Context
Up to 256K
Use Case
RAG / Enterprise
command-arerank-3.5embed-v4
Rerank / Embed v4 / STTNon-commercial only
Access: dashboard.cohere.com → API Keys → Trial key. Resets monthly. Not for commercial production use.
Note on Cline: Cline itself is a free, open source coding agent extension, not a model host. It has no free tokens of its own; every request is billed through whichever provider key you connect (BYOK), except for a small number of models tagged FREE inside its own usage-billing option. It is listed under Coding Tools below rather than as a free token source.
What Cursor, Codex, and Antigravity Actually Give You Free
The three coding environments founders ask about most
| Tool | Free Tier | What's Included | Paid Trigger |
|---|---|---|---|
| Cursor | Hobby plan, $0 | 2,000 Tab code completions/month, limited Agent requests, free trial of Pro | Pro is $20/mo, includes $20 of frontier model usage billed at raw provider rates after that pool |
| OpenAI Codex | Bundled with ChatGPT Free and Go plans | Codex extension works in VS Code with a ChatGPT login; free accounts get reduced usage caps vs. Plus, Pro, Business, Enterprise | Higher limits and priority require ChatGPT Plus ($20/mo) or above |
| Google Antigravity | Free with a Google account | Google's agent-first coding environment, ships with Gemini 3.5 Flash access out of the box since the May 2026 launch | No separate paywall reported yet; usage tied to same Google AI quotas as AI Studio |
| Cline (VS Code extension) | Extension is free forever, open source | No tokens of its own. Bring your own key (BYOK) from any provider in this guide, or use its built-in usage-billing option where a small number of models are tagged FREE | Usage is billed by whichever provider key you connect, not by Cline itself |
Practical takeaway: If you already code in VS Code, install Cline first and connect it to a free key from this guide, such as OpenRouter or NVIDIA NIM. Cursor and Antigravity are full IDEs with their own free tiers layered on top, worth trying separately for a more polished agent-first UI.
The OpenRouter Free Models Router
One endpoint, zero cost, automatic model selection
OpenRouter ships a special model slug called openrouter/free that removes the need to pick a specific free model yourself.
How It Works
Setup
Set model to openrouter/free in any OpenAI-compatible request. No other code changes needed.
Selection
The router randomly picks one of the ~29 available free models per request, filtering automatically for whichever capability your request needs, such as vision, tool calling, or structured outputs.
Context
200,000 tokens on the router endpoint itself, since it must work across every model it might route to.
Rate Limits
Same as any other free model: 50 requests/day by default, up to 1,000/day once you hold a $10+ credit balance, and 20 RPM across all users.
Trade-off
You do not know which model answers each request, so output style and quality vary between calls. Some free models on the router also have known tool-calling failures. Treat it as a discovery and testing tool, not a stable production endpoint.
The Recommended Free Stack
How to cover 80-90% of your agent workload at $0/month
Agent Workflow Stack (July 2026)
Default Route
Google AI Studio (Gemini 3.5 Flash) for general chat, document Q&A, and multimodal tasks. Rate-limited but the strongest free multimodal option available.
Hermes / Agent Runtime
NVIDIA NIM for running Hermes Agent, OpenClaw, and Open Code against GLM-5.2, now a permanently free tier with no credit balance to track. Watch the 40 RPM ceiling on complex multi-step tool calls, since that is the only real constraint left.
Real-Time/Voice
Groq for latency-sensitive flows: voice agents, streaming chat. 300+ tokens/sec plus free Whisper STT.
Batch Speed
Cerebras when wall-clock generation speed, not daily request count, is your bottleneck.
RAG Pipeline
Cohere trial key for Rerank 3.5 and Embed v4. Best-in-class retrieval at zero cost, capped at 1,000 calls/month.
Overflow / Frontier
DeepSeek pay-as-you-go for the 10-20% of calls needing more than free tiers allow, still priced far below GPT and Claude per-token rates.
Low-Cost Subscriptions for Unlimited Open Weights
What you actually pay, what it includes, and the top models in each plan
These are flat-fee subscriptions, not pay-per-token pricing. Use these once you outgrow free tier rate limits and need guaranteed, repeatable throughput for a running agent.
GLM Coding Lite
$18
Featherless Basic
$10
Featherless Premium
$25
GLM Coding Pro
$72
Featherless Agent Std
$100
GLM Coding Max
$160
Featherless Max
$200
| Provider / Plan | Price | What's Included | Top 3 Open Weight Models (July 2026) | Best For |
|---|---|---|---|---|
| Featherless Basic | $10/mo | Unlimited monthly requests, 2 concurrent connections, private/anonymous usage, no logs | Llama 3.3 8B/70B, Mistral 7B, Qwen2.5 14B | Chat, role-play, lightweight assistants |
| Featherless Premium | $25/mo | Unlimited requests, 4 concurrent connections, full catalogue access | Kimi K2, GLM-5.2, DeepSeek V4 Flash | Founders who want one flat fee for any open model, any size |
| Featherless Agent Standard/Max | $100-$200/mo | Up to 8 concurrent connections, 256K context, one agent sandbox included | DeepSeek V4 Pro, GLM-5.2, Nemotron 3 Ultra | Running persistent agents (Hermes Agent, OpenClaw) at scale |
| GLM Coding Lite | $18/mo (~$12.60/mo annual) | ~120 prompts per 5-hour cycle, flagship GLM access via Claude Code, Cline, or Roo Code | GLM-4.5, GLM-5.2 (flagship), GLM-5.2-Air | Day-to-day coding on small to mid repos |
| GLM Coding Pro | $72/mo (~$50.40/mo annual) | ~5x Lite usage, priority queueing | GLM-5.2, GLM-5.2-Air, GLM-4.5 | Heavier individual development workloads |
| GLM Coding Max | $160/mo (~$112/mo annual) | ~20x Lite usage, peak-hour priority | GLM-5.2, GLM-5.2-Air, GLM-4.5 | Teams or power users running agents continuously |
| Hugging Face PRO | $9/mo | Removes cold starts, higher inference limits, ZeroGPU Spaces access | Qwen2.5-72B, Llama 3.3 70B, FLUX.1-dev | Prototyping without cold-start delays |
Specialty Models: Image, Video, TTS/STT
Free access beyond text generation
| Modality | Provider | Free Model | Free Limit | Access |
|---|---|---|---|---|
| Image Generation | Hugging Face | FLUX.1 Dev | Credit-based, shared infra | HF Inference API, read token |
| Image Generation | Cloudflare Workers AI | Stable Diffusion XL | 10K neurons/day | Workers AI API |
| Speech-to-Text | Groq | Whisper Large v3 / Turbo | 2,000 req/day, 7,200 audio-sec/min | console.groq.com audio endpoint |
| Speech-to-Text | Cloudflare Workers AI | Whisper | Included in 10K neurons/day | Workers AI API |
| Speech-to-Text | Cohere | Audio Transcriptions | 5 RPM (trial key) | Cohere trial key |
| Vision / Multimodal | Google AI Studio | Gemini 3.5 Flash | Rate-limited, no hard cap | ai.google.dev |
| Vision / Multimodal | OpenRouter | Select :free vision models | 20 RPM / 50-1,000 RPD | openrouter.ai |
| Embeddings | Cohere | Embed v4 | Included in 1,000 calls/mo trial | Cohere trial key |
| Embeddings | Cloudflare Workers AI | BGE / Multilingual embeddings | 10K neurons/day | Workers AI API |
| Reranking (RAG) | Cohere | Rerank 3.5 | Included in 1,000 calls/mo trial | Cohere trial key |
| Video Generation | Z.ai / GLM | CogVideoX-3 (paid, low cost) | Not free, $0.20/video | docs.z.ai |
Note: Free video generation is essentially nonexistent as of July 2026. Z.ai/GLM video models are extremely cheap ($0.20-$0.40 per video) but not free. If your project needs video, budget for this separately.
Benchmarks: Top Free Models Right Now
5 leading open weight models plus the top free Gemini models, verified July 2026
| Model | Type | Top Coding Score | Top Reasoning Score | Top Agentic Score | Context Window | Best Used For |
|---|---|---|---|---|---|---|
| GLM-5.2 (Z.ai) | Open Weight, MIT | 62.1% SWE-bench Pro | 80.3% GPQA Diamond | 81.0% Terminal-Bench 2.1 | 1,000,000 tokens | Long-horizon coding across large monorepos; strongest open model on context-heavy agent work |
| DeepSeek-R1 | Open Weight, MIT | ~49% SWE-bench Verified | Strong on AIME/GPQA-class reasoning chains | Competent multi-step tool use | 128,000 tokens | Budget-friendly deep reasoning and math-heavy tasks where speed matters less than accuracy |
| Qwen3 Coder 480B | Open Weight, Apache 2.0 | Strong on SWE-bench and LiveCodeBench-class evals | Solid general reasoning, not a frontier leader | Reliable tool-calling in agent harnesses | 256,000 tokens | Dedicated coding assistant use inside IDEs and CI pipelines |
| Kimi K2 (Moonshot AI) | Open Weight, Modified MIT | Competitive on agentic coding benchmarks | Strong general knowledge and instruction-following | Built specifically for tool-use and agent workflows | 128,000 tokens | General-purpose agent backbone when GLM-5.2 rate limits are exhausted |
| Llama 4 Scout (Meta) | Open Weight, Llama License | Moderate, below coding-specialist models | Balanced general reasoning | Native multimodal, works well as a fast agent router | 10,000,000 tokens (theoretical); most hosts cap at 128K-1M | High-throughput classification, summarization, and vision tasks at low latency |
| Gemini 3.5 Flash (Google) | Closed, Free Tier | Reportedly beats Gemini 3.1 Pro on most coding evals | Frontier-class reasoning at Flash speed | Strong native agentic and tool-use performance | 1,048,576 tokens | The best all-around free model for coding agents, multimodal input, and everyday reasoning |
| Gemini 3 Flash-Lite (Google) | Closed, Free Tier | Improved over Gemini 2.5 Flash-Lite, below full Flash | Good for its size, not frontier-class | Basic tool use, best for simple agent loops | 1,048,576 tokens | High-volume, latency-sensitive tasks where you need the highest daily free request cap |
Reading the table: GLM-5.2 currently leads every other open weight model on raw coding and long-context agentic tasks, which is why it anchors both the NVIDIA NIM and OpenRouter free tiers in this guide. Gemini 3.5 Flash is the strongest closed, fully free model available, largely because Google subsidizes it to build developer habit ahead of Vertex AI billing.
Fallback Benchmarks: Low-Cost Frontier Models
Where to go when free limits run out, compared on cost, coding, reasoning, and context
| Model | Price (In/Out per 1M) | SWE-bench Verified | SWE-bench Pro | Top Reasoning/Agentic | Context Window | Best Used For |
|---|---|---|---|---|---|---|
| Gemini 3.5 Flash | Lowest-cost Flash tier pricing | Beats Gemini 3.1 Pro on most evals | Not separately reported | Frontier-level agentic and tool use at Flash speed | 1,048,576 tokens | Best cost-to-capability ratio for coding agents and multimodal work |
| Gemini 3 Flash | $0.50 / $3.00 | Outperforms Gemini 2.5 Pro | Not separately reported | Near-Pro-level reasoning at Flash cost | 1,000,000 tokens | Everyday business tasks, drafts, summarization at the lowest reliable frontier cost |
| Claude Sonnet 5 | $2.00 / $10.00 (intro through Aug 31, 2026; $3/$15 standard after) | 85.2% | 63.2% | 80.4% Terminal-Bench 2.1, 81.2% OSWorld-Verified | 1,000,000 tokens (128K standard output, up to 300K via beta header) | Highest coding accuracy per dollar among named frontier models; best default for serious agent coding |
| Grok 4.5 | $2.00 / $6.00 (under 200K context; 2x above) | Not separately reported | 64.7% | 83.3% Terminal-Bench 2.1, 62.0% DeepSWE 1.0 | 500,000 tokens | Token-efficient coding agent; uses roughly 4x fewer output tokens than Opus-class models on equivalent tasks, lowering real-world cost |
| GPT-5.6 Terra | $2.50 / $15.00 | Matches GPT-5.5-class quality per OpenAI | Not separately reported | Balanced everyday reasoning and tool use | 1,050,000 tokens | Middle-tier fallback for teams standardized on the OpenAI ecosystem needing reliable general performance |
| GPT-5.6 Luna | $1.00 / $6.00 | Below Terra and Sol, above GPT-5 Mini | Not separately reported | Fast, lightweight agent tasks | 400,000 tokens | Cheapest OpenAI fallback for high-volume, low-complexity agent calls |
How to read this: Claude Sonnet 5 currently posts the highest SWE-bench Verified score of this group at 85.2 percent, but its introductory pricing expires August 31, 2026, after which it rises to $3/$15. Grok 4.5 is the value play for agent-heavy workloads since its token efficiency can offset its output price. Gemini's Flash tiers remain the cheapest entry point into frontier-adjacent performance for any team not tied to a specific ecosystem.
5 Rules for Staying Inside Free Limits
1. Cache repeated answers. Most rate limits count every call, not unique queries. A cached response is worth hundreds of API calls saved.
2. Right-size your model. Classification and summarization do not need a 70B model. Save the largest free models for reasoning and generation steps.
3. Trim your context aggressively. Since most limits measure tokens, a lean 400-token system prompt beats a 2,000-token one every time.
4. Route by task type. Interactive chat to Groq, bulk processing to Cerebras, multimodal to Google AI Studio, Hermes-style agents to NVIDIA NIM. Stacked limits multiply your effective free quota.
5. Watch RPM before RPD. Most 429 errors come from bursting requests too fast, not from hitting a daily ceiling. Add small delays between agent tool calls to stay under the per-minute limit.
Disclosure: Free tier limits and model rosters change without warning. Google reduced Gemini free quotas in late 2025. NVIDIA NIM removed its earlier 5,000-credit trial ceiling in 2026 and now runs a permanently free tier gated only by a 40 RPM hard cap; its terms still permit using free-tier inputs and outputs to improve their models. OpenRouter's higher daily request cap (1,000 RPD vs. 50 RPD) depends on maintaining a $10+ funded balance, not usage history. Always confirm current RPM, RPD, and per-model caps on the provider's own docs before routing business-critical or client data through a free API. This guide is republished on the 5th of each month.
Monthly guide
Keep your AI stack current without doing the research twice.
PromptHacker turns the fast-moving AI market into clear decisions, useful workflows, and practical next steps.