See what you pay, what context you get, and where the best value lives for coding, writing, and high-volume usage.
If you want the shortest pricing answer, start with Mistral Small 3.1 for the best value default. Use Mistral: Mistral Nemo only when raw lowest API price matters more than output quality.
Cheap does not automatically mean efficient. The real pricing decision is whether lower token cost saves more money than the extra review, rewrites, or mistakes it creates.
This page compares raw cost, context, and practical usefulness so you can avoid false-economy pricing decisions.
The safest value pick, the raw cheapest API, and the fast default worth considering before you optimize around price alone.
This table focuses on the pricing decisions teams actually make first: best value default, absolute cheapest option, budget coding pick, and a fast low-cost option.
Ultra-cheap multimodal model for massive-volume, low-complexity pipelines.
Use this section to decide whether you should optimize for raw API cost, value per prompt, cheaper coding throughput, or faster user-facing response time.
Ultra-cheap multimodal model for massive-volume, low-complexity pipelines.
Ultra-high-volume classification, summarisation, and lightweight vision tasks
You need reliable multi-step reasoning or coding quality — it won't hold up.
Pricing recommendations are based on a mix of list price, real-world usefulness, speed, context window, and whether a lower-cost model still holds up under practical workloads.
Benchmark scores from SWE-bench (coding), ARC-AGI-2 (reasoning), and MMLU (knowledge breadth) — cross-referenced against Chatbot Arena community votes to filter out cherry-picked provider claims.
Input and output costs verified directly against each provider's official API pricing page. Updated whenever a price change is detected. Value-per-dollar is weighted separately from raw benchmark rank.
Advertised context sizes are noted but scored against real-world usability — models that degrade significantly at large contexts are penalised even if the window is technically available.
Production signals matter more than lab scores. We weight Cursor and Windsurf defaults, HackerNews sentiment, developer surveys, and which models teams actually keep using after the honeymoon period.
One-off wins on cherry-picked benchmarks don't move our rankings. We favour models that stay dependable across repeated prompts, diverse task types, and long sessions without degrading.
Time-to-first-token and output throughput from Artificial Analysis speed benchmarks. Latency is categorised from Very fast to Deliberate — relevant when building interactive or high-throughput products.
Data sources
See your monthly API cost vs consumer subscription across all models.
600 input + 700 output tokens
| Model | Monthly API cost | Annual API cost | vs Subscription |
|---|---|---|---|
MistralMistral Small 3.1 | $0.40 | $4.86 | API only |
OpenAIGPT-4o Mini | $0.76 | $9.18 | API cheaper Sub wins at 39,216 msg/mo |
DeepSeekDeepSeek V3 | $1.40 | $16.78 | API only |
MetaLlama 4 Scout | $1.71 | $20.52 | Free via Meta AI |
MetaLlama 4 Maverick | $2.22 | $26.64 | Free via Meta AI |
DeepSeekDeepSeek R1 | $2.79 | $33.53 | API only |
API costs are estimates based on the token counts above and listed per-million-token prices from each provider. Subscription plans include usage caps and may not cover all models — check provider pages for current limits. Prices update from our database when providers change their rates.
Compare 119 models by cost profile, provider, and context.
Mistral Nemo is open-weight (Apache 2.0 license), so self-hosting is an option for teams that want to eliminate API costs entirely. Pricing via API is through Mistral's La Plateforme. The model uses a Tekken tokenizer which is more efficient than older Mistral tokenizers, especially for non-English text.
Being open-weight, this model can be run locally or self-hosted via providers like Together AI, Fireworks, or Groq, often at even lower costs. The 16K context window is a meaningful limitation compared to other models in this price tier.
As an open-weight model, Llama 3 8B can be self-hosted via platforms like Ollama, Replicate, or Together AI. The 8,192 token context window is a significant practical limitation. Pricing listed reflects hosted API inference; self-hosted costs vary.
Pricing reflects API access through third-party providers; Google also offers Gemma 2 9B weights for free download and self-hosting. The 8,192 token limit is a hard architectural constraint of this version.
Pricing is exceptionally competitive at $0.05/$0.08 per 1M tokens. Available via Mistral's La Plateforme API and various third-party providers. GDPR-friendly EU-based hosting is a notable advantage for European enterprise customers. No image input or output support.
Priced at a flat $0.10/1M for both input and output, making cost estimation predictable. The '2512' suffix indicates a December 2025 release version. Best suited for batch processing, classification, or extraction pipelines where volume is high and task complexity is low.
Output cost of ~$0.20/1M tokens is notably higher relative to input cost — factor this in for verbose generation tasks. Best suited for inference pipelines where outputs are short and structured. Available via multiple inference providers due to open-weight licensing.
Mistral Small 3.2 is available as an open-weight model, making it deployable on-premises or via self-hosted infrastructure — a key differentiator over GPT-4o Mini and Claude Haiku for privacy-sensitive use cases.
The '8B 2512' in the model name likely refers to a specific versioned release; despite the naming, this is based on Mistral's 3B architecture. Confirm parameter count and capabilities with Mistral's official documentation before production use.
This is v0.1, the original release — not to be confused with v0.2 or v0.3 which substantially improve context length and quality. The listed context window of ~2,824 tokens is unusually small even among budget models. Marked as superseding Mistral Large 2 in the spec, which appears to be a data error — this model does not supersede Mistral Large 2 in capability or positioning.
Llama Guard 4 supports the MLCommons hazard taxonomy and is designed to be used as a shield model in multi-model architectures. Not suitable as a standalone AI assistant. Available via Meta's open model ecosystem and third-party API providers.
Pricing is among the lowest available in any major provider's lineup as of mid-2025. Context window of 1M tokens is a significant differentiator at this price tier. Check Google AI Studio and Vertex AI for rate limits on high-volume usage.
This is an open-weights safety/moderation-specific model, not a general assistant. Pricing reflects its budget-tier positioning. Availability may be limited or subject to change as it appears to be a research/infrastructure model rather than a consumer product. Verify OpenAI's terms around usage and redistribution for the OSS weights.
Worth considering for internal search, analysis, and review workflows where data sovereignty matters.
Available via Mistral API and can be self-hosted via open weights. Pricing is among the lowest available for a code-specialized model. Designed to work within coding agent frameworks like SWE-agent and OpenHands.
Model name suggests a December 2025 revision ('2512'). Pricing is symmetric at $0.20/1M for both input and output, which simplifies cost modeling. Confirm availability on your target API platform as Mistral model availability varies by provider.
Context window of 32,768 tokens is notably smaller than competing budget models. Pricing is approximate ($0.10 input / $0.30 output per 1M tokens). Availability is through Mistral's API (La Plateforme) and may also be accessible via third-party providers. Confirm fine-tune scope before deploying for non-creative tasks.
Voxtral Small is audio-in capable but does not support image input. The 32K context window is notably short for a 2025 model. Pricing is via Mistral's API; availability through third-party providers may vary. Check whether your use case requires audio input — the text-only version of Mistral Small 3.1 may be more appropriate for pure text workloads.
Output cost of ~$0.40/1M tokens means output-heavy workloads (long generations) will accumulate cost faster than input-heavy ones. Best suited for tasks where outputs are short-to-medium length. No image generation capability.
Available via multiple inference providers including Together AI, Fireworks, and OpenRouter. As an open-weight model, it can also be self-hosted for even lower marginal costs at scale. Part of Meta's Llama 3.2 family which also includes a 90B vision variant for heavier workloads.
Pricing listed is for standard (non-cached) input/output. Context caching is available and can reduce costs significantly for repeated long-context calls. Image and audio inputs are priced separately. Free tier available via Google AI Studio.
Pricing is approximate based on listed rates. As a 'Lite' model, it may not support all multimodal features available in full Flash or Pro variants. Check Google AI Studio for feature availability and rate limits.
This is a preview model (09-2025 versioned) and may be subject to breaking changes or deprecation. Pricing is approximate based on listed rates. Not recommended for production systems requiring SLA guarantees. Check Google AI Studio or Vertex AI for GA alternatives.
Pricing is $0.10/1M input and $0.40/1M output tokens. Officially supersedes GPT-4o in OpenAI's lineup for lightweight use cases. Context window of ~1.047M tokens is one of the largest available at this price tier.
This model is designed exclusively for content moderation and safety classification tasks. It follows the MLCommons AI Safety benchmark taxonomy. It should be deployed as a guardrail layer alongside generative models, not as a replacement for them. Not suitable for end-user-facing conversational applications.
As an open-weight model, Gemma 4 26B can also be self-hosted, making API pricing largely irrelevant at scale. The 'A4B' suffix denotes the active parameter count in its MoE configuration. Listed as superseding Gemini 3 Flash Preview, though Gemini 2.0 Flash remains a stronger hosted alternative.
As an open-weight model, Gemma 4 31B can be self-hosted via Ollama or Hugging Face in addition to Google's API. Pricing shown is for hosted inference. No image input capability confirmed at launch.
GPT-4o Mini punches well above its price for classification, summarisation, and simple writing. It struggles when tasks get complex.
Strong strategic fit for teams thinking about data sovereignty or custom fine-tuning.
Pricing at $0.15/$0.60 per million tokens makes this one of the most affordable capable models on the market. Available via Mistral's La Plateforme API and compatible with OpenAI-style endpoints. No image input support confirmed at launch.
Pricing shown is via third-party API providers (e.g., OpenRouter, Together AI) — costs may vary. Meta releases Llama 3.1 weights publicly, enabling self-hosting at even lower cost. Not available directly from Meta as a hosted API.
Pricing reflects Mistral API rates and may vary by reseller. The model's name 'Saba' references Arabic linguistic heritage, signaling its intended multilingual focus. No vision or tool-use capabilities documented at launch.
Pricing is highly competitive at $0.30 input / $0.50 output per million tokens. Context window is 131K tokens. No vision/image input support. xAI's API platform is newer and may have availability or rate-limit considerations compared to established providers.
Model is in Beta — API behavior, rate limits, and availability may change without notice. No multimodal support confirmed. Reasoning mode may increase effective latency on complex prompts despite fast base speed.
At $0.10/1M input, the cost question disappears. The only question is whether the task complexity exceeds what Mistral Small can handle.
Pricing is symmetric at $0.54/1M for both input and output. As an open-weight model, costs can drop significantly if self-hosted. The 32K context window is a hard ceiling — plan accordingly for document-heavy workflows.
Available via Mistral's La Plateforme API. Also accessible through Continue.dev, Cursor, and other IDE integrations that support the Codestral endpoint. FIM (fill-in-the-middle) mode is specifically supported for autocomplete use cases. Output price rounds to ~$0.90/1M tokens.
This is the original Llama 3 70B, not the 3.1 or 3.3 variants. Llama 3.1 70B offers a 128K context window at comparable pricing and is strongly preferred. Consider this model only if you have a specific reason to pin to the original Llama 3 checkpoint.
Symmetric input/output pricing at $0.65/1M tokens is straightforward but positions it oddly — it's pricier than GPT-4o Mini while lacking its multimodal features. Available via multiple inference providers including Google Vertex AI and third-party APIs.
DeepSeek V3 shocked the market on release. At this price point with this capability level, it forces a reconsideration of when premium models are actually worth it.
Claude 3 Haiku is part of the original Claude 3 family (March 2024). Anthropic has since released Claude 3.5 Haiku, which is generally recommended over this model for new use cases. Still widely available via Anthropic API and AWS Bedrock.
Pricing is asymmetric: input at ~$0.20/1M is excellent, but $1.50/1M output undercuts its budget appeal for generation-heavy use. Availability through xAI's API; check for rate limits and regional availability as xAI's infrastructure is still scaling.
The default budget pick for startups watching cost. The 1M context at this price is unmatched.
Pricing shown is $0.40 input / $1.60 output per 1M tokens. Cached input tokens are significantly cheaper. The 1M token context window is a standout feature at this price tier — few competitors match it. Supersedes GPT-4o as the recommended default for cost-conscious applications.
Pricing of $0.50 input / $1.50 output per 1M tokens places it firmly in the budget-flagship category. Available via Mistral API (La Plateforme) and major cloud providers. December 2025 update ('2512') improves instruction following over the earlier 2407 release.
GPT-3.5 Turbo is still available via OpenAI API and supports fine-tuning, which keeps it relevant for teams with existing trained models. However, OpenAI has deprioritized its development in favor of the GPT-4o family. Not multimodal — text only.
Output cost of $2/1M tokens is higher than some competing budget models (Gemini Flash at ~$0.60/1M output). At scale, output-heavy tasks may erode cost advantages — monitor token ratios carefully. Supersedes GPT-4o, which may be deprecated on a rolling basis.
At $2/1M output tokens, costs can accumulate in verbose code-generation tasks — monitor output token usage carefully in agentic loops. Not a general-purpose flagship replacement; best deployed alongside a stronger model for planning/reasoning layers.
The December 2025 (2512) release date suggests this is a recent iteration. Pricing at $0.40 input / $2.00 output is notably competitive for a code-specialist model with 256K context. Verify availability and rate limits via Mistral API or partner providers.
Pricing is notably aggressive at ~$0.40 input / $2.00 output per 1M tokens. Available via Mistral's La Plateforme API. Part of the Devstral family, which is distinct from Mistral's general-purpose Mistral Medium line.
Priced at $0.40 input / $2.00 output per 1M tokens. Officially supersedes Mistral Large 2, making it an easy drop-in upgrade for existing Mistral users. Available via Mistral's API and La Plateforme.
Officially supersedes Mistral Large 2, representing a generational shift in Mistral's lineup toward multimodal capability at lower cost tiers. Available via Mistral API and select cloud providers. No function calling limitations noted at this tier.
R1 is a genuine milestone for open-source AI. The reasoning quality is real — the tradeoff is latency, not capability.
Output cost ($2.5/1M) is disproportionately higher than input cost ($0.3/1M), so generation-heavy use cases may see costs add up faster than expected. Thinking/reasoning mode may be available but incurs additional cost.
The 32,768 token context window is unusually small even for a budget model — verify this limit hasn't changed before deploying in production. The 'Nano Banana' name appears to be an internal or experimental identifier; confirm model availability and stability via Google AI Studio or Vertex AI before relying on it in critical workflows.
Audio tokens are priced differently from text tokens in OpenAI's API — audio input/output carries a significant premium over text tokens, so real-world costs for voice-heavy workloads will be substantially higher than the listed text token price suggests. Check OpenAI's audio token pricing separately.
This is a pinned legacy snapshot (v0613) and may eventually be deprecated by OpenAI. The 4,095-token context window is its most significant practical limitation. OpenAI's own GPT-4o mini offers drastically more context and better quality at a comparable price — strongly consider migrating.
This is a preview model and may have limited availability, unstable rate limits, and pricing that changes before general availability. Output cost at $3/1M is notably higher than input cost, so applications generating long outputs should budget accordingly.
Uses the legacy /v1/completions endpoint, not /v1/chat/completions. The 4,095-token context window is a hard constraint that makes it unsuitable for most modern tasks. OpenAI has not deprecated it, but it receives no capability updates.
Ideal for teams running thousands of daily coding prompts where premium model costs add up quickly.
Output cost of $2/1M tokens is unusual — lower than input cost, which favors use cases with long inputs but short outputs like image captioning or document summarization. Verify image generation token pricing separately, as image outputs are often billed differently by OpenAI.
Output cost of $4/1M is notably higher than competing fast/mini models. Input cost at ~$0.80/1M is competitive. Best value emerges in input-heavy pipelines like document classification or RAG retrieval where output tokens are minimal.
Great for drafts, rewrites, and quick-turn internal workflows where Anthropic's tone quality matters.
Supports three reasoning effort settings via the API (low, medium, high), which significantly affect latency and token usage. No vision/image input support. Available via OpenAI API and ChatGPT Plus.
The 'High' suffix refers to the reasoning_effort parameter set to 'high', which increases token usage and latency significantly versus o3 Mini at medium or low effort. Priced at $1.1/$4.4 per million tokens, it is far cheaper than o1 ($15/$60) and full o3, making it attractive for batch workloads.
Priced at $1.1/$4.4 per 1M tokens (input/output), o4 Mini is significantly cheaper than o3 ($10/$40) and o4. Output tokens are 4x the input price, so verbose reasoning traces can add up — use max_completion_tokens limits in production pipelines.
The 'High' suffix denotes maximum reasoning effort, distinct from o4 Mini (balanced) and o4 Mini Low. Higher effort means higher token consumption in internal reasoning traces, which can push effective cost above the stated $1.1/$4.4 per million for very complex queries. No image generation capability.
Priced at $1/1M input and $5/1M output tokens, placing it above true budget models like Gemini Flash but below mid-tier flagships. Confirm availability of extended thinking or tool-use features via Anthropic's API documentation, as Haiku-tier models sometimes receive these capabilities later than Sonnet/Opus.
Best when you specifically need an OpenAI model in your stack.
OpenAI has been gradually deprecating older GPT-3.5 variants. Availability may be limited or sunset in the future. At $3/$4 per million tokens, this is dramatically overpriced relative to its capability in 2024-2025.
Best when you want near-flagship coding quality with a massive context window at a mid-tier price.
The EU hosting angle is the real differentiator here — for teams outside Europe, other models perform better.
Available via Mistral API and as open weights (Apache 2.0 license) for self-hosting. The open-weight option is a key differentiator for privacy-sensitive or on-premise deployments. API pricing at $2/$6 per million tokens is mid-range but faces pressure from newer, cheaper alternatives.
Available via Mistral API (la Plateforme) and supports self-hosted deployment. The '2411' suffix indicates a November 2024 release. Supersedes Mistral Large 2 as the primary flagship. Image input pricing follows the same $2/1M token rate.
Priced at $2/1M input and $8/1M output tokens — cheaper than GPT-4o at launch. The 1M context window is real but performance near the ceiling is less tested than Gemini's equivalent. No built-in image generation or voice modality.
Pricing at $2/$8 per 1M input/output tokens is moderate for a reasoning model, but long internal reasoning traces can significantly inflate output token counts. Not available via all API tiers — check OpenAI access levels.
Deep Research mode requires agentic tool access (web browsing); pricing reflects token usage but research tasks can consume significant tokens across multi-step retrieval loops. Availability may depend on API tier or organizational access level. Not a drop-in replacement for the standard o4 Mini in general-purpose workflows.
Pricing shown is for prompts under 200K tokens; inputs over 200K tokens are billed at $2.50/1M input and $15/1M output. Gemini 2.5 Pro includes built-in 'thinking' (reasoning) mode which can increase latency and cost further.
This is a preview model (05-06 date suffix indicates a versioned snapshot); Google may deprecate or change it without long notice. Confirm production readiness before building critical pipelines on this endpoint. The 1M context window applies to text and multimodal inputs combined.
This is a preview model (06-05 date suffix indicates a versioned snapshot); Google may deprecate or modify it before a stable GA release. Pricing tiers differ based on prompt length — prompts over 200K tokens are charged at $2.50/1M input and $15/1M output, significantly increasing cost for very long-context use cases.
Pricing is asymmetric: cheap on input ($1.25/1M) but expensive on output ($10/1M), so it favors read-heavy or summarization tasks over verbose generation. The 400K context window is one of the largest available at this price tier. Supersedes GPT-4o, which remains available at lower cost for lighter workloads.
Pricing is asymmetric — input is relatively affordable at $1.25/1M but output at $10/1M can accumulate quickly in agentic or verbose-output workflows. Cached input pricing may apply through the OpenAI API. Not to be confused with GPT-5 reasoning variants (o-series) which use chain-of-thought and have separate pricing.
The $10/1M output cost means heavy code generation workloads can get expensive fast — budget carefully for bulk generation use cases. Context window of 400K is among the largest in its price tier. Supersedes GPT-4o, so existing GPT-4o coding workflows should consider migrating for improved performance.
Pricing structure heavily favors input-heavy use cases like RAG and retrieval. The $10/1M output cost makes it expensive for long-form generation at scale. Context window of 400K is competitive but not best-in-class against Gemini 3.1 Pro's 1M+ window.
Output cost of $10/1M tokens is asymmetric compared to the $1.25 input price — high-volume generation tasks will become expensive quickly. No vision or image generation confirmed based on available specs. Supersedes GPT-4o in the OpenAI lineup but does not replace o-series reasoning models.
Asymmetric pricing ($1.25 input / $10 output) rewards read-heavy workflows like code review and repo analysis over generation-heavy tasks. The 400K context window is among the largest in the balanced price tier. No image input/output support confirmed at launch.
Output cost of $10/1M tokens is the key budget consideration — input is competitively priced but output costs mirror GPT-4 Turbo-tier pricing. Best paired with a cheaper model for lightweight or repetitive coding subtasks. Context window of 400K is well-suited to monorepo analysis but verify token limits on your deployment tier.
Strong when your work lives between visuals, messaging, and product context.
Audio tokens are counted differently from text tokens — a few seconds of audio can consume hundreds of tokens, so monitor usage carefully. Real-time audio streaming requires WebSocket or Realtime API endpoints, not the standard Chat Completions API. Availability may be limited by tier or region.
The 2M context window is a genuine competitive advantage — no other frontier model gets close for document-heavy workflows.
This is a preview model — API behavior, pricing, and availability may change before general release. The 65K context window is unusually constrained for a Gemini Pro-tier model; double-check if your use case requires longer contexts before committing.
Output cost of $14/1M tokens is the primary budget consideration — workloads with high output-to-input ratios will accumulate costs quickly. No image generation capability. Supersedes GPT-5.2, which should be deprecated or deprioritized.
Priced asymmetrically with low input cost ($1.75/1M) and high output cost ($14/1M), which rewards concise prompting but penalizes verbose code generation. The 400K context window is one of the largest available at this price tier. Supersedes GPT-5.2 with improved multi-file coherence; users on GPT-5.2 should migrate. No multimodal input support confirmed at launch.
Thinking tokens (the internal reasoning trace) count toward output token billing, which can significantly increase costs on complex queries. The thinking budget can often be configured via the API. Best used selectively for tasks that genuinely benefit from deliberation rather than as a default model.
Pricing at $3 input / $15 output positions this as a 'balanced' tier model, but output costs are notably higher than comparable models like GPT-4o ($10 output). Extended context (200K) is available by default. Check Anthropic's API for rate limits and availability by tier.
Supersedes Claude 4 Haiku, positioning it as a step-up option rather than a true budget model. The 1M token context window is the headline feature. Output cost of $15/1M tokens is on the higher end for this tier — compare to Gemini 3.1 Pro at roughly $10/1M output before committing to high-volume use.
Powers Cursor and Windsurf by default. If your team already uses either, you're already using this model.
Available via xAI API and integrated into X Premium subscriptions. Real-time X data access is a differentiating feature not available on competing models. Pricing is competitive but output costs are on the higher end for balanced-tier models.
Model is currently in beta, meaning capabilities and pricing may change. Real-time X data integration depends on xAI's API access policies, which may be subject to change. No image generation support confirmed.
Flat $10/1M input and output pricing is unusual — most flagship models charge more for output tokens. Verify whether image token costs (typically higher per effective token) are included under this pricing or billed separately, as OpenAI historically charges additional fees for image inputs.
Unique value is the computer-use capability. If you're building agents that operate software, nothing else compares right now.
Pricing is $5 input / $25 output per 1M tokens — identical output cost to GPT-5.4 tier models. Note the 'Supersedes Claude 4 Haiku' label appears to be a data anomaly; Opus 4.5 is the top-tier model, not a Haiku replacement. Confirm model availability on the Anthropic API dashboard as Opus-tier models sometimes have access restrictions.
Best reserved for complex multi-file refactors, architecture decisions, and agentic coding pipelines where mistakes are expensive.
Ranked from public benchmark and pricing data verified April 26, 2026: SWE-Bench Pro 64.3%, 1M context, $5/$25 per 1M tokens.
Pricing at $6 input / $30 output per million tokens is significantly higher than GPT-4o ($2.50/$10). Best accessed via Anthropic API or Amazon Bedrock. Claude 3.5 Sonnet (October 2024 version) supersedes the June 2024 release with improved performance.
GPT-4 Turbo is available via the OpenAI API. It has largely been succeeded by GPT-4o, which is faster, supports vision natively, and is cheaper. Organizations should evaluate whether migrating to GPT-4o or o3 makes more sense before building new workflows on this model.
This is a pinned model snapshot (v1106) and will not receive capability updates. OpenAI may deprecate older snapshots over time. Knowledge cutoff is April 2023. Not recommended for new deployments given the superior cost-performance of GPT-4o and GPT-4.1.
This is a 'preview' variant that OpenAI has largely deprecated in favor of gpt-4-turbo and gpt-4o. The endpoint may be retired or redirected by OpenAI without notice. Check the OpenAI model deprecation schedule before building production applications on this model.
Deep Research mode involves agentic tool calls and web browsing, which can multiply effective token costs significantly. Pricing is per token but real-world research sessions often consume large amounts of both. Available via ChatGPT Plus/Pro and API; API access may require higher usage tiers.
At $15 input / $60 output per 1M tokens, a single complex back-and-forth session can cost dollars. o1-mini is available at a fraction of the price for lighter reasoning tasks. OpenAI has since released o3 and o3-mini, which largely supersede o1 for most reasoning use cases.
Output pricing at $75/1M tokens is among the highest in the market — nearly 3x GPT-4.1's output cost. Batch API discounts may be available through Anthropic. Context window is 200K but very long prompts at Opus pricing can become extremely expensive quickly. Note: supersedes field lists Claude 4 Haiku, which is likely a data error — Opus 4.1 more logically succeeds Claude Opus 4.
At $30/$60 per million tokens, this is one of the most expensive text-only models available. The 8,191-token context window is a hard ceiling that makes it unsuitable for most document-processing tasks. OpenAI continues to offer it for API backward compatibility but actively recommends migrating to GPT-4o or GPT-4 Turbo. New projects should not default to this model.
This is a frozen March 2023 snapshot of GPT-4, not a current model. OpenAI may deprecate legacy snapshots with limited notice. The 8,191-token context window is a hard constraint. Cost is identical to much more capable current models, making this a poor choice for new projects.
o3 Pro is only available via the OpenAI API and ChatGPT Pro subscription tier. Response times can range from tens of seconds to several minutes depending on problem complexity. Output pricing at $80/M tokens is 4x the cost of standard o3.
Output cost of $120/1M tokens is exceptionally high and will compound quickly in agentic or multi-turn workflows. Budget carefully. Context window of 400K is generous but falls short of Gemini 3.1 Pro's 1M+ offering for ultra-long document tasks.
At $15 input / $75 output per 1M tokens, Opus 4 is one of the most expensive models available. Anthropic recommends using Claude Sonnet 4 for most production use cases and reserving Opus 4 for tasks explicitly requiring maximum capability.
Worth considering only if you have existing integrations built around this model.
Ranked from public benchmark and pricing data verified April 26, 2026: SWE-Bench Pro 58.6%, Terminal-Bench 2.0 82.7%, $5/$30 per 1M tokens, 1M API context.
o1-pro is available only via the OpenAI API and ChatGPT Pro subscription ($200/month). It does not support streaming and has longer latency than any other OpenAI model. Not suitable for high-volume workloads.
Reserved for future partners around monitoring, optimization, procurement, and evaluation.
The AI-native editor most developers switch to when they want GPT-4 and Claude working inside their actual codebase — not a chat window next to it.
The fastest way to get a sourced, current answer to any question. Pairs well with longer-form AI tools — use it to verify, then use Claude or GPT to synthesize.
Newsletter
Get concise updates when input costs, output costs, or value rankings change.
No spam. Useful updates only. Affiliate disclosures always clearly labeled.
Mistral: Mistral Nemo is the cheapest raw API option in the current directory, but Mistral Small 3.1 is the better cheap default for most teams.
Mistral Small 3.1 is the best cheap AI API here because it balances low cost, high speed, and broad usefulness better than the absolute cheapest options.
Pay for a premium model when quality failures create expensive rework, missed edge cases, or costly downstream mistakes. Premium models rarely make sense for low-stakes high-volume prompts.
Meta: Llama 3.1 8B Instruct is the strongest budget coding specialist in the directory, while Mistral Small 3.1 is the better low-cost generalist if the work extends beyond pure coding.
Mistral Small 3.1 is the best price-to-usefulness default for most teams.
Mistral: Mistral Nemo is the lowest-cost option by list price, but it is not automatically the best low-cost decision.
Anthropic: Claude 3 Haiku is the better pick when low latency matters almost as much as low spend.
A dirt-cheap multilingual model perfect for bulk text tasks, but don't expect frontier-level reasoning.
The right tool for cheap, fast, high-volume tasks — not for anything that requires serious thinking.
A capable budget workhorse, but Claude 3.5 Haiku has made it mostly obsolete for new deployments.
A dirt-cheap multilingual model perfect for bulk text tasks, but don't expect frontier-level reasoning.
Teams needing a cheap, fast, multilingual workhorse for classification, summarization, or light coding tasks at scale.
You need reliable multi-step reasoning, advanced code generation, or any image/multimodal processing.
The right tool for cheap, fast, high-volume tasks — not for anything that requires serious thinking.
High-throughput applications where cost and speed matter more than frontier-level quality, such as chatbots, content classification, and text summarization.
You need deep reasoning, long document analysis, complex code generation, or outputs where quality directly impacts user trust.
A capable budget workhorse, but Claude 3.5 Haiku has made it mostly obsolete for new deployments.
High-volume production pipelines, customer support bots, and real-time text processing where cost and latency are critical constraints.
You need deep reasoning, complex coding tasks, or high-quality creative writing — Claude 3 Sonnet, GPT-4o Mini, or even Claude 3.5 Haiku will serve you better.
One API key to access GPT-5, Claude 4, Gemini, Llama, and 100+ other models. Ideal for developers who want to switch models without rewriting integration code.
These tools are independently recommended based on real-world fit with the models on this site. Links may include affiliate or referral tracking — see our disclosures.
