AI Model Costs on magicdoor.ai: Current Chat Rates

Kimi update — October 28, 2026: Choose Kimi K3.1 Pro for the current Kimi route through Moonshot AI via Vercel AI Gateway. K3 and Renewable K3 are retired. See the Kimi workflow and limits guide before continuing an older Kimi chat.

Direct answer: The model rate is only one part of your cost. First choose a supported model, estimate the input and output tokens for your actual task, then add any tool charges and the plan. The table below shows current USD per million tokens for the chat picker, not a quality ranking or a guaranteed monthly bill. Use the model-selection guide to choose a model for your task, and this guide to compare price units and limits.

DeepSeek checkpoint update — October 21, 2026: Current DeepSeek labels, links and any DeepSeek prices on this page now refer to V4.1 Flash 1017, using the October 21 launch source packet. Earlier fact-check and source-access dates below apply to the original comparison or study, not this checkpoint. Other vendor claims and historical measurements were not reverified in this update.

Fact-checked October 9, 2026 against the merged Haiku 5.5 registry, Anthropic Haiku pricing, and Mistral Large 4 model reference plus Vercel AI Gateway endpoint metadata. Haiku 5.5 supports vision and PDFs, but has no hosted code interpreter. The Mistral entry is a public preview whose availability depends on the current picker and Gateway endpoint; check the current model index and picker before use. Rates can change, especially the Mistral sale price.

Chat picker rates at a glance

These are the models in the merged picker registry, not all historical chat aliases. The live picker remains the availability check. Above 100K input tokens, Haiku charges $0.50 input, $0.05 eligible cached input, and $2.50 output per million for the whole request; five-minute cache writes use $0.125 / $0.625 up to / over 100K. For example, uncached 10,000 input + 2,000 output tokens cost $0.002; 120,000 input + 2,000 output cost $0.065. Compare Haiku and GPT-6 Luna before a long task. Auto is an Experimental setting that selects among a smaller answer pool, not an additional separately priced model; usage is billed for the concrete answer model. Provider-specific cache writes, request-based tools, and long-context multipliers are not included in the three displayed columns. Consult the model detail pages for those conditions.

Make a comparable estimate

For a request with 100,000 new input tokens, 50,000 eligible cached input tokens, and 10,000 output tokens, the Large 4 October 8 sale rates produce $0.0924 in model usage ($0.068 new input + $0.0035 cache reads + $0.0209 output). This illustrative request stays below Magicdoor's 262,144-token input budget, but cache eligibility is not guaranteed. The quote excludes subscription and separately priced tools.

For a smaller example, 10,000 new input tokens and 1,000 output tokens at the same sale rates cost $0.00889 before other charges. The Gateway lists up to 262,144 output tokens but that is a maximum, not a suggested completion length. Mistral's direct model documentation lists 1M context; Magicdoor's Gateway route does not expose that full limit. The Large 4 Preview guide explains the route and privacy scope.

Image generation prices

Image generation is priced per image, not per chat token. The selected model and quality tier determine the quote. Active image generation costs $0.012–$0.50 per image; Recraft Upscaler is a separate $0.006 operation.

At the $1 renewal usage allowance, that is roughly 2–83 images before other usage, depending on the selected generator and quality. Basic's first month includes $2 usage, not $1; purchased top-ups are separate and do not expire. Read the image models page and live quote before generation.

Subscription and other charges

Basic and Plus plans provide the same supported models and features. Basic is $6/month with $2 included usage balance in month one, then $1 on renewal. Plus is $12/month with $7 monthly usage balance. Included monthly balance resets; separately purchased top-ups do not expire. You pay for usage above the balance. A plan price is not an unlimited-token allowance, and historical $8–10 monthly-spend estimates do not establish your bill.

Image generation/editing is charged per image or operation, not per chat token; compare current image tools before sending. Search and other paid tools can add charges beyond chat tokens. Long conversations re-send context; start a new thread when a prior topic is no longer needed. For cost-sensitive work compare the same measured token mix on a second model, then evaluate the answer quality yourself rather than ranking models by rate alone.

Open chat to compare a real task after checking current model pricing and availability.

FAQ

Does Mistral Large 4 have a permanent price?

No. On October 8, 2026, its pinned Vercel AI Gateway endpoint lists sale rates of $0.68 per million input tokens, $0.07 per million eligible cached input tokens, and $2.09 per million output tokens. The Magicdoor route fails closed if its guarded rates change until pricing is updated. Check current model pricing before use.

How much does Claude Haiku 5.5 cost?

Haiku 5.5 costs $0.10 input and $0.50 output per million tokens for requests with up to 100,000 input tokens. Above that threshold, the whole request uses $0.50 input and $2.50 output rates, with separate cache pricing.

What counts as cached input?

Only input tokens reported as eligible cache reads receive a model's cached-input rate. New input is charged at the input rate. Cache-write rates and long-context tiers may also apply to other models; do not assume an entire prompt is cached.

Is the subscription included in token prices?

No. Basic is $6 per month and includes $2 usage balance in the first month, then $1 on renewal; Plus is $12 with $7 monthly usage balance. Additional use is metered. Purchased top-ups do not expire, while monthly included balance resets.

Can I compare prices using one blended figure?

Use the same measured workload and compare input, eligible cached input, output, and any tool charges separately. A blended figure depends on the chosen token mix and can hide tier thresholds or sale changes.