AI and automation

Which AI model should run a small business job every day?

When an AI job runs every day, you stop buying a chat subscription and start buying tokens. A token is roughly three quarters of a word, and the price per million tokens is the number that decides what a daily report or a nightly CRM sweep costs. This table compares each vendor's workhorse model, the one priced for volume rather than the flagship.

Checked against each vendor's own pages on October 14, 2026. Every cell links to its source. Prices and plans change often, so click through before you buy.

Claude

Anthropic

Workhorse model
Claude Sonnet 5.5Mid-tier by price ($2/$10 versus Opus 5.5 at $4/$20 and Fable 5.1 at $10/$50). The pricing page calls Opus 5.5 the 'Daily driver for agentic coding and enterprise work' and Sonnet 5.5 'High-performance model for coding and agents', so Sonnet 5.5 is a cost-based pick for everyday volume. Model ID claude-sonnet-5-5, released September 28, 2026.Sonnet 5.5 model page Claude pricing page
API price per million tokens
$2 input, $10 output per million tokens (standard API)Batch API is a 50% discount on input and output. Cache read $0.10 / MTok, 5m cache write $2.50 / MTok, 1h cache write $4 / MTok. No future effective date shown.Sonnet 5.5 model page Claude pricing page
Context window
1M tokens (max output 128K tokens)Batch API can reach 300K output tokens with the output-300k-2026-03-24 beta header.Sonnet 5.5 model page
Knowledge cutoff
Jun 2026 (reliable knowledge cutoff and training data cutoff)Both rows read Jun 2026 in the Capabilities table.Sonnet 5.5 model page
API data retention
Not retained by default for Sonnet 5.5; zero data retention on requestAnthropic's API retention page says prompts and outputs are not retained by default, except for the Fable and Mythos models, which require 30-day retention. Zero data retention is enabled per organization through sales.API and data retention Commercial data retention policy
Structured output
Yes. JSON schema output through output_config.format, generally available.GA on the Claude API, Claude Platform on AWS, Google Cloud and Microsoft Foundry. On Amazon Bedrock it is GA only on the legacy Bedrock integration for Opus 4.6 and earlier, not on Claude in Amazon Bedrock. Unsupported schema features return a 400 error: recursive schemas, numerical and string length constraints, array constraints beyond minItems 0 or 1, external $ref, and additionalProperties other than false.Structured outputs docs
Where else you can run it
Amazon Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS (also the Claude API directly)On-premises is not listed. Model IDs differ on Bedrock (anthropic.claude-sonnet-5-5); the other platforms use claude-sonnet-5-5. Bedrock structured outputs have a caveat noted in that row.Sonnet 5.5 model page
Our pick
The default for daily jobs that touch customer data: Sonnet 5.5 is not retained by default, returns JSON on a schema, and runs on three clouds.

GPT

OpenAI

Workhorse model
GPT-6.1 SolModels page: GPT-6 Astra is the most capable flagship, GPT-6 Luna the most efficient for high-volume tasks, GPT-6.1 Sol sits between.OpenAI models page OpenAI GPT-6.1 Sol model page
API price per million tokens
$2.00 input, $10.00 output per 1M tokens (standard, short context)Short context means 272K input tokens or less. Long context (over 272K) is $4.00 input, $15.00 output. Cached input is $0.10. Batch is $1.00 / $5.00. Regional processing endpoints carry a 10% uplift for models released on or after March 5, 2026.OpenAI pricing page OpenAI GPT-6.1 Sol model page
Context window
1,050,000 tokens (1.05M)Max output is 128,000 tokens.OpenAI GPT-6.1 Sol model page
Knowledge cutoff
April 30, 2026OpenAI GPT-6.1 Sol model page
API data retention
Abuse monitoring logs kept up to 30 days by default. Zero Data Retention is offered, subject to prior OpenAI approval.The page marks many endpoints as ZDR eligible (chat/completions, responses, embeddings, images, audio, moderations, realtime and others), some with limitations. Removed the unconfirmed exclusion claim.OpenAI data controls guide
Structured output
Yes, JSON schema structured outputs in the APIThe guide says Structured Outputs is available in the latest models starting with GPT-4o and recommends gpt-6-astra for new projects. It does not name GPT-6.1 Sol, so model-specific support is not stated on that page.OpenAI structured outputs guide
Where else you can run it
Not publishedNeither the model page nor the models page lists Azure, Vertex or Bedrock availability for GPT-6.1 Sol. Bedrock appears only in navigation and Codex/agents links, Azure and Google Cloud only under workload identity federation. Not a listing of the model.OpenAI models page OpenAI GPT-6.1 Sol model page
Our pick
Same price as Claude at the workhorse tier with a 1M context. Pick it when the team already lives in ChatGPT.

Gemini

Google

Workhorse model
Gemini 3.8 Flash (gemini-3.8-flash)Mid-tier is a judgment call: the models page does not use that label. It calls 3.8 Flash the most intelligent Flash model, listed as Stable, and the pricing page lists it alongside Gemini 4 Argon (flagship), 3.6 Flash and 3.5 Flash-Lite.Gemini API models page Gemini API pricing
API price per million tokens
Input $0.75, output $3.75 per 1M tokens, Standard paid, through Dec 31, 2026Rises to $1.50 input and $7.50 output starting January 1, 2027. Output includes thinking tokens.Gemini API pricing
Context window
1,048,576 input tokens, 65,536 output tokensInput and output token limits as shown on the model page.Gemini 3.8 Flash model page
Knowledge cutoff
Not publishedNeither the model page nor the Google Cloud 3.8 Flash page states a knowledge cutoff. The model page shows only a Latest update of September 2026, which is a page date.Gemini 3.8 Flash model page
API data retention
Paid: no set period given; logged for abuse checks. Zero retention not listedPaid Services logs are kept for a limited period solely to detect and prevent violations, with no stated duration. Grounding with Google Search stores prompts and output for 30 days. Unpaid services content may be used to improve Google products.Gemini API terms
Structured output
Yes, JSON Schema via response_format, a subset of the specSupports a subset of JSON Schema; very large or deeply nested schemas may be rejected (the nesting claim was not re-confirmed on the page).Gemini API structured output docs
Where else you can run it
Google Cloud: a Gemini 3.8 Flash page exists on Gemini Enterprise Agent Platform. AWS Bedrock, Microsoft Azure and on-premises are not listedThe Google Cloud page title is Gemini 3.8 Flash | Gemini Enterprise Agent Platform, but the readable text does not list platforms (page was truncated; last 18k characters unread). The Gemini API pages list no cloud platforms. Redirect from cloud.google.com to docs.cloud.google.com is within google.com.Gemini 3.8 Flash on Google Cloud Gemini API models page
Our pick
Cheapest workhorse here at under a dollar per million in. The promotional price has an end date, so check it.

Muse

Meta

Workhorse model
muse-spark-1.3 (Standard mode). Page does not label it mid-tier.Muse Wonder (muse-wonder-1.0) is the most capable model, priced $4.00/$20.00. Muse Spark is positioned for high-volume, latency-sensitive or cost-sensitive work, and 1.3 is 'Recommended for new work'. No page calls any model 'mid-tier'.Models - Meta Model API
API price per million tokens
$1.25 input / $4.25 output per 1M tokens (Standard mode)Cached input $0.15 per 1M. Contributor mode is $0.10 input / $0.20 output, and Meta may use that content to train models. The pricing page shows no effective date. Figures come from /docs/pricing-rate-limits.Pricing and rate limits - Meta Model API
Context window
1,048,576 tokens (about 1M)Models page lists a 1,048,576-token context window for Muse Spark versions; the muse-spark model page says 1M.Models - Meta Model API Muse Spark 1.3 - Meta AI
Knowledge cutoff
Not publishedNeither the Models page, the model page nor the docs home states a knowledge cutoff.Models - Meta Model API
API data retention
Retention period not published. Standard Services content is not used to train Meta models. Zero data retention is not described on the commitments page.Commitments page says retention is 'as described in the Terms and Privacy Policy' and gives no period. The pricing page mentions ZDR only in the Muse Voice Transcribe section (priced at parity with Standard), not for Muse Spark. Contributor mode content may be used for training.Meta Model API Data Commitments Pricing and rate limits - Meta Model API
Structured output
Yes. JSON schema through response_format with type json_schema.Output is constrained at decode time.Structured output - Meta Model API
Where else you can run it
No AWS Bedrock, Vertex AI or Azure listed. Vendor page names OpenRouter.OpenRouter is a third-party router. No on-premises option for Muse Spark is stated. Muse Glimmer is open weight and self-hosted (vLLM, SGLang, llama.cpp, ExecuTorch), but that is a separate model. The Models page does not mention AWS, Azure, Vertex or OpenRouter.Muse Spark 1.3 - Meta AI Models - Meta Model API
Our pick
Second cheapest after Gemini, with a 1M context, and Standard mode does not train on your data. Skip the discounted Contributor mode for business work: that one does.

Mistral

Mistral AI

Workhorse model
Mistral Medium 3.5Listed in the Flagship models table on the API pricing page.Mistral API pricing
API price per million tokens
$1.5 input, $7.5 output per million tokens (cached input $0.15)Price is on docs.mistral.ai/inference/pricing, not mistral.ai/pricing. No effective date shown.Mistral API pricing Medium 3.5 model card
Context window
256k tokensThe model card shows "256k" as shown by the vendor.Medium 3.5 model docs
Knowledge cutoff
Not publishedNeither the model page nor the model card states a knowledge cutoff.Medium 3.5 model docs
API data retention
Default retention period not stated. Zero data retention is available on request, at Mistral's discretion, pay-as-you-go and stateless API calls only.Stateful products (Vibe, Chat, agents, libraries, Files API) are excluded from ZDR. Free-mode users must first enable pay-as-you-go.Can I activate ZDR Data retention policy
Structured output
Yes. Built-in JSON mode plus custom structured outputs; the model card lists Structured Outputs under features.The docs page does not use the phrase JSON schema; it says custom structured outputs. The model card lists Structured Outputs via /v1/chat/completions and /v1/conversations.Structured output docs Medium 3.5 model card
Where else you can run it
Open weights under a Modified MIT license (self-deploy). Mistral lists Azure AI, Amazon Bedrock, Google Cloud Vertex AI, Snowflake Cortex, IBM watsonx and Outscale for its models in general; the Medium 3.5 pages do not name any cloud.The deployment page does not mention Medium 3.5 specifically, so cloud availability for this exact model is not confirmed.Mistral deployment Medium 3.5 model docs
Our pick
Open weights you can run yourself, with zero data retention available. The option when the model must stay in-house.

Price the job before the model

A small business job is small. The cheapest workhorse model is rarely the deciding factor. These three usually are.

  1. 01Count tokens firstA daily summary of 50 emails at roughly 800 tokens each is about 40,000 tokens in and 2,000 out. On every price row here that is around ten cents a day.
  2. 02Retention matters moreIf the job touches customer data, the API retention policy and the option for zero retention outweigh a few cents of price.
  3. 03Structured output saves hoursA model that returns clean JSON on request plugs into a spreadsheet or CRM without cleanup code. That is the feature that makes automation stick.

Pick one vendor, write the job down in four lines, and run it for a month before comparing anything else. Switching models is cheap. Switching halfway through a half-built job is not.

Spearworks AI

Have a job that should run itself every morning?

I build small daily AI jobs for owner-operated businesses: pull, summarize, draft, file. You review the output; the job does the typing. Fixed quote after one conversation about the job.

Book a free consult

Nathan Spear · spearworksai.com

Spearworks Labs · All comparisons · Checked October 14, 2026. Not affiliated with any vendor listed.