Four Claudes, One Budget: Which Claude Model to Use for Every Performance Marketing Job

Anthropic's newest models differ by 100 times in price, and the right pick depends on the job, not the leaderboard.

Tiger Tracks · Eye of the Tiger · AI and Automation · October 2026

Between September 1 and October 7, 2026, Anthropic released four new Claude models [2]. The most expensive, Claude Fable 5.1, lists at $50 per million output tokens; the cheapest, Claude Haiku 5.5, starts at $0.50 [1]. A token is a small chunk of text, often part of a word, that AI providers use to meter and bill usage. In the illustrative example later in this article, that 100-to-1 gap means one weekly job costs about $2 or $225, depending on which model runs it.

That gap is a budget decision hiding inside a dropdown menu. Marketing work is also token-heavy: in Anthropic's June 2026 usage study, conversations mapped to marketing managers used about 2.5 times as many tokens as those mapped to editors [11]. This article explains what each Claude model is for, which performance marketing jobs fit each one, and how to stop paying premium rates for routine work.

1. Anthropic Now Sells Four Tiers of Claude, and the Top One Costs 100 Times the Bottom One

Anthropic's current lineup has four models, each built for a different kind of work [1]. All four read text and images, and all four accept up to 1 million tokens of input at once [1]. That limit, called the context window, is roughly the amount of material the model can hold in view during one task. The differences are speed, depth and price. "Agentic" work, which appears in Anthropic's descriptions below, means the model takes a series of actions on its own, such as searching, writing code or editing files.

Anthropic's default advice is to start with Opus 5.5 for most workloads [1]. It also says Opus 5.5 performs at the level of Fable 5.1 on most work while costing 40% less to run than its predecessor, Opus 5, at default settings on typical workloads [3]. That is a vendor claim, but the flagship is no longer the automatic choice.

2. The Best Model Is the Cheapest One That Clears Your Quality Bar

The most common misreading is that the most powerful model is always the safest choice. The evidence points the other way for many tasks. On GPQA Diamond, a set of graduate-level science questions, Sonnet 5.5 alone tied Opus 5.5 alone at 91%, and Haiku 5.5 scored 85% [7]. The tie partly reflects Opus refusing more biology questions; on questions both answered, Opus led by about 2 points [7]. Sonnet costs half as much per token [1], so that edge is worth paying for only when the task needs it.

Paying more does not always buy more, either. In an internal coding test, pairing Opus 5.5 with Fable 5.1 as an adviser added 1.7 points for about 2.1 times the cost. Anthropic describes that gain as at the edge of run-to-run noise [7].

Cheaper per token is also not the same as cheaper per task. Bigger models often think longer, and that thinking is billed. DeepLearning.AI's newsletter The Batch reported that Fable 5.1's cost per task rose about 20% over Fable 5, despite a 75% cut to its cached input price, the rate for text the model has already processed and can reuse [13]. The price list is fact; the lesson is that task cost, not token price, is the number to manage.

3. Haiku 5.5 Handles the High-Volume Plumbing of Performance Marketing

Performance marketing runs on repetitive classification. Search terms need intent labels. Product feed rows need clean attributes. Reviews and comments need sentiment tags. Anthropic built Haiku 5.5 for this kind of work: summaries, classification and database queries at high volume [4]. It says Haiku 5.5 costs around 75% less to run than Haiku 4.5 on average [4].

Worked example: tagging 50,000 search terms

Consider a weekly job that labels 50,000 search terms as brand, competitor, high-intent or low-intent. The numbers below are illustrative. Assume each term needs about 300 input tokens (instructions plus the term) and returns about 30 output tokens. The Batch API (application programming interface), Anthropic's half-price option for jobs that can wait, is shown in the last column [9][10].

1. Input: 50,000 terms × 300 tokens = 15 million tokens.
2. Output: 50,000 terms × 30 tokens = 1.5 million tokens.
3. Haiku 5.5: 15 × $0.10 plus 1.5 × $0.50 = about $2.25 [1].
4. Fable 5.1: 15 × $10 plus 1.5 × $50 = $225 [1].

Inputs: 15 million input and 1.5 million output tokens per run, list prices [1], and Haiku's entry rate for prompts up to 100,000 tokens [4]. Batch API discount of 50% as listed for Opus 5.5, Sonnet 5.5 [9][10] and Fable 5.1 [5], and assumed for Haiku 5.5. Excludes thinking tokens, the tokens a model spends reasoning before it answers, which add cost.

A weekly tagging job does not need results in seconds, so batch is the obvious setting. Over a year, the gap between the cheapest and most expensive choice is about $117 versus $11,700 at list prices without batch (TT est.).

4. Sonnet 5.5 Is the Everyday Workhorse for Copy, Analysis and Scripts

Most day-to-day marketing work sits in the middle: drafting ad and email variants, summarizing a week of channel data, writing a Google Ads script or a tag manager fix. Anthropic lists code generation, data analysis and content creation as Sonnet 5.5's core use cases [6]. It is fast and costs half of Opus [1].

Sonnet fits when a person reviews the output anyway, as a copywriter does with headlines. The human is the quality layer, so extra model depth adds cost, not safety.

5. Opus 5.5 Earns Its Price on Judgment Calls and Long Builds

Some jobs reward depth: a cross-channel audit that must reconcile platform data with analytics, a quarterly business review (QBR) narrative, a media mix model (MMM) readout a chief financial officer will question, or a landing page build that runs for hours. Anthropic recommends Opus 5.5 for complex agentic coding and enterprise work, including multihour autonomous coding and vision-heavy workflows [6].

Customer quotes in the launch announcement point to writing quality. Ramp's John Ruelas said, "It writes like a good colleague, and follows our writing rules" [3]. Those are vendor-selected testimonials, so treat them as signals, not proof. Opus 5.5 is also available with zero data retention, an arrangement under which the provider does not store prompts or outputs after responding [3]. That matters when client data is in the prompt.

6. Fable 5.1 Is a Specialist for the Longest, Hardest Jobs, With a Data Trade-Off

Fable 5.1 is Anthropic's most capable model open to all customers [6]. Anthropic suggests it for agent sessions that run for hours, multistep deep research, and analysis carried through to a finished document, spreadsheet or deck [6]. In marketing terms, that means a full category and competitor study for a pitch, built from a blank page.

The trade-offs are real. Per token it costs 2.5 times Opus 5.5 [1]. Fable models carry 30-day data retention and are not available under zero data retention unless Anthropic expressly authorizes it [5]. In the Claude apps, Fable runs only on paid plans, and Pro and standard Team seats pay for it with usage credits from the start [14]. Anthropic's guidance is to use Fable for demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort (a setting covered in section 7) still falls short [1].

7. Effort Settings and Model Pairing Cut Cost Without Switching Models

Every current Claude model has an effort setting that controls how much it thinks before answering, from low through medium, high, extra-high and max [8]. Anthropic states that "tuning effort is often a better lever than switching models" [6]. Opus 5.5 and Haiku 5.5 default to medium; the others default to high [8].

The practical result can be large. In the Opus 5.5 launch, Carl Bennett, chief information officer of Deloitte Consulting, reported: "On US consulting analysis, low thinking effort matched its higher thinking settings on half the output and passed our quality checks" [3]. Half the output means roughly half the output tokens billed, for output that passed the same checks.

Pairing models is the other lever. In an orchestrator setup, a frontier model plans the work and hands the heavy reading to cheaper workers. On one large test, that setup cost 47% to 55% less than Fable 5.1 alone but scored 10 to 12 points lower [7]. On another, the earlier Claude Fable 5 alone matched the coordinator setup's accuracy at 22% to 30% lower cost [7]. The saving is real, and so is the quality cost, which is why pairing needs testing on your own work.

8. Model Routing Ripples Into Budgets, Governance and Team Roles

The first-order effect is lower cost for routine AI work. Routing, sending each task to the cheapest model that can do it, turns the price gap into savings.

The second-order effect is volume. When tagging 50,000 terms costs about a dollar (TT est.), teams run it daily instead of monthly, and more AI output reaches campaigns. More output means more to check. Adoption is already broad. Nearly nine in ten respondents to McKinsey's 2026 survey report regular use of AI in at least one business function [12]. About one in ten say costs have constrained their organizations' use of AI agents [12].

The third-order effect is governance. Model choice becomes part of the client data agreement, because retention terms differ by model [3][5]. AI spend starts to look like a media line, with someone accountable for which model handles which job.

9. Three Futures for Model Choice in Marketing

The most likely future holds because the release cadence shows no sign of slowing and each tier still has a distinct price [1][2]. The wild card is data terms. If retention rules keep differing by model, client contracts, not quality, may decide which model a team can use [3][5]. The strategic implication is the same in every scenario: teams that test models on their own work and track cost per task will spend less and trust their output more.

10. Recommendations by Audience

For Marketing and Brand Leaders

• Set a default model per job type, and write down when a task earns a more expensive model.
• Track cost per task, not price per token.
• Check retention terms before any client or customer data goes into a prompt.

For Marketing Operations and Analytics Teams

• Run each recurring job on two tiers for a week and compare accuracy against a human-checked sample.
• Use the Batch API for any job that can wait.
• Test lower effort settings before switching to a cheaper model.

For Technology Providers

• Show model, effort and cost per task in your product's reporting.
• Let customers set routing rules and spend caps by job type.

Questions Readers Ask

Which Claude model should a marketing team use by default?

Anthropic recommends starting with Opus 5.5 for most workloads [1]. For high-volume, simple jobs, Anthropic suggests starting with Haiku 5.5, testing thoroughly, and upgrading only for specific capability gaps [6].

Is Fable 5.1 worth the price for marketing work?

For multi-hour research that ends in a finished deck or model, it can be. For routine copy and reporting, it costs 2.5 times Opus 5.5 per token without a clear payoff [1].

Can my team use Fable 5.1 in the Claude app?

Only on paid plans. Max and premium Team seats include some Fable usage; Pro and standard Team seats pay with usage credits [14].

What is the effort setting?

It controls how long the model thinks before answering. Lower effort is faster and cheaper; higher effort helps on hard reasoning [8].

Conclusion

Start this week with one recurring job, such as search term tagging or the weekly performance summary. Run it on two tiers, compare the output against a human-checked sample, and keep the cheaper model if it passes. Then write the rule down. Models will keep changing; a habit of testing cost per task will outlast every release. The machine does the volume. People set the quality bar and decide what earns the premium. That is the Human-Led, AI-Augmented approach.

References

1. Anthropic. (Accessed October 2026). Models overview (Claude Platform documentation). https://platform.claude.com/docs/en/models/overview
2. Anthropic. (Accessed October 2026). Release notes (Claude Help Center). https://support.claude.com/en/articles/12138966-release-notes
3. Anthropic. (September 22, 2026). Introducing Claude Opus 5.5. https://www.anthropic.com/claude-opus-5-5
4. Anthropic. (October 7, 2026). Introducing Claude Haiku 5.5. https://www.anthropic.com/claude-haiku-5-5
5. Anthropic. (Accessed October 2026). What's new in Claude Fable 5.1 (Claude Platform documentation). https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1
6. Anthropic. (Accessed October 2026). Choosing a model (Claude Platform documentation). https://platform.claude.com/docs/en/about-claude/models/choosing-a-model
7. Anthropic. (Accessed October 2026). Optimizing for cost and intelligence (Claude Platform documentation). https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence
8. Anthropic. (Accessed October 2026). Effort (Claude Platform documentation). https://platform.claude.com/docs/en/build-with-claude/effort
9. Anthropic. (Accessed October 2026). Claude Opus 5.5 (Claude Platform documentation). https://platform.claude.com/docs/en/models/opus-5-5/overview
10. Anthropic. (Accessed October 2026). Claude Sonnet 5.5 (Claude Platform documentation). https://platform.claude.com/docs/en/models/sonnet-5-5/overview
11. Anthropic. (June 26, 2026). Anthropic Economic Index report: Cadences. https://www.anthropic.com/research/economic-index-june-2026-report
12. McKinsey & Company. (August 25, 2026). The state of AI in 2026: On the road to ROI. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
13. DeepLearning.AI. (September 11, 2026). Fable Holds The Top Spot (For Now). The Batch. https://www.deeplearning.ai/the-batch/fable-holds-the-top-spot-for-now
14. Anthropic. (September 2, 2026). Claude Fable models on your plan (Claude Help Center). https://support.claude.com/en/articles/15424964-claude-fable-models-on-your-plan

Published by Tiger Tracks. Eye of the Tiger Intelligence Series.