DEV Community: Emcy

Claude Fable 5 vs Mythos 5 vs Opus vs Sonnet vs Haiku: Which Claude Model Should You Use?

Emcy — Wed, 10 Jun 2026 08:50:02 +0000

Originally published at https://codeculture.store/blogs/developer-culture/claude-fable-5-vs-mythos-opus-sonnet-haiku.

Claude Fable 5 vs Mythos 5 vs Opus vs Sonnet vs Haiku: Which Claude Model Should You Use?

Claude's model lineup now has a new top end. As of June 10, 2026, the practical hierarchy is Fable 5 and Mythos 5 at the top, Opus 4.8 below them for complex work, Sonnet 4.6 as the balanced daily-driver model, and Haiku 4.5 as the fastest low-cost option.

The shortest recommendation: use Fable 5 for the hardest generally allowed work, Mythos 5 only if you have approved trusted access, Opus 4.8 for high-end reasoning with fewer Fable-specific constraints, Sonnet 4.6 for everyday serious work, and Haiku 4.5 for speed and cost.

Key Takeaways

Fable 5 is Anthropic's most capable widely released model.

Mythos 5 uses the same underlying model as Fable 5, but is restricted to approved users.

Opus 4.8 remains the strong complex-reasoning option below the Mythos-class tier.

Sonnet 4.6 is the best default for balanced price, speed, and intelligence.

Haiku 4.5 is the fast, inexpensive choice for simpler tasks and high-volume workflows.

Quick Decision Table

If you need...	Use this model
Maximum broadly available Claude capability	Claude Fable 5
Approved advanced cyberdefense or trusted research access	Claude Mythos 5
Complex reasoning or agentic coding without jumping to Fable	Claude Opus 4.8
Strong daily coding, writing, analysis, and agent work	Claude Sonnet 4.6
Fast and cheap responses at scale	Claude Haiku 4.5

Sources: Anthropic Fable/Mythos announcement and Claude models overview.

Model Specs At A Glance

Model	Best for	Access	Context	Max output	Input / output price
Claude Fable 5	Hardest general reasoning, coding, vision, long-horizon work	Generally available	1M	128k	$10 / $50 per MTok
Claude Mythos 5	Approved trusted cyberdefense and research workflows	Limited availability	1M	128k	$10 / $50 per MTok
Claude Opus 4.8	Complex reasoning and agentic coding	Available	1M on most surfaces	64k	$5 / $25 per MTok
Claude Sonnet 4.6	Balanced speed and intelligence	Available	1M	64k	$3 / $15 per MTok
Claude Haiku 4.5	Fastest low-cost near-frontier work	Available	200k	64k	$1 / $5 per MTok

Note: Anthropic's docs list Microsoft Foundry context differences for Opus 4.8 and platform-specific details. Always verify against the provider surface you are using before migration. Source: Claude models overview.

Claude Fable 5: The New Hard-Task Default

Claude Fable 5 is the model to test when the task is too complex, too long, or too ambiguous for your normal Claude workflow.

Use Fable 5 for:

large codebase migrations;
long-running agentic coding tasks;
deep research synthesis;
complex document, chart, and table analysis;
high-value vision tasks;
work where failure costs more than model spend;
tasks where Sonnet or Opus need too many correction loops.

Fable 5 is not automatically the right model for every prompt. It is expensive relative to Sonnet and Haiku, and it introduces stricter safeguards around some sensitive domains. Anthropic also documents 30-day retention for Fable 5 API traffic, so enterprise teams need policy approval before sending sensitive workloads. Source: Claude Platform release notes.

Use Fable when the extra capability changes the outcome, not when it merely makes a simple answer slightly nicer.

Claude Mythos 5: The Controlled-Access Twin

Claude Mythos 5 is not a normal upgrade path. It shares the same underlying model as Fable 5, but is offered only to approved users, initially through Project Glasswing and trusted access programs.

Use Mythos 5 only if:

your organization has approved access;
the task falls inside an allowed trusted-use program;
you need capabilities Fable 5 deliberately restricts;
your team can handle the policy, monitoring, and compliance expectations around that access.

For almost everyone else, Mythos 5 is best understood as the reason Fable 5 has safeguards. It is the same capability tier, but in a different access and risk environment.

Claude Opus 4.8: The High-End Workhorse Below Fable

Claude Opus 4.8 remains important because not every serious task needs Fable 5. Anthropic describes Opus 4.8 as its most capable Opus-tier model for complex reasoning, long-horizon agentic coding, and high-autonomy work. Source: Claude models overview.

Use Opus 4.8 for:

complex coding work that does not need Fable;
high-stakes reasoning where Sonnet is not enough;
long-context work where cost still matters;
tasks affected by Fable-specific safeguards;
fallback paths when Fable refuses or routes.

Opus also costs half as much as Fable 5 on listed input and output pricing. If Opus solves the job reliably, it is the more economical model.

Claude Sonnet 4.6: The Balanced Daily Driver

Claude Sonnet 4.6 is the model most teams should try first for routine serious work. Anthropic describes Sonnet 4.6 as the best combination of speed and intelligence, with a 1M token context window and lower cost than Opus or Fable.

Use Sonnet 4.6 for:

daily coding help;
content drafts;
product analysis;
support workflows;
moderate refactors;
internal tools;
agent tasks where latency and price matter.

Sonnet is the default when you do not yet know that the task needs a frontier model. In practice, a strong workflow is to start with Sonnet, escalate to Opus for difficult work, then use Fable only for the hardest or highest-value tasks.

Claude Haiku 4.5: Fast, Cheap, And Good Enough

Claude Haiku 4.5 is the volume model. It is the fastest and cheapest model in this comparison, while still being positioned by Anthropic as having near-frontier intelligence.

Use Haiku 4.5 for:

classifications;
short summaries;
extraction;
routing;
formatting;
simple customer support workflows;
high-volume transformations;
first-pass drafts that will be reviewed by another model or a human.

The mistake is using Fable for work Haiku can do. If the task is low-risk, short-context, and easy to verify, Haiku is often the smarter business choice.

Workflow Recommendation For Developers

For AI coding and developer workflows, think in escalation tiers:

Workflow stage	Recommended model
Cheap first pass, extraction, routing	Haiku 4.5
Daily implementation help	Sonnet 4.6
Hard bug, refactor, or architecture reasoning	Opus 4.8
Long-horizon migration or high-value agent task	Fable 5
Approved trusted cyber/research work	Mythos 5

This is how you avoid model maximalism. The strongest model is not always the best model. The best model is the cheapest one that reliably clears the quality bar.

Claude Shirts For The Model-Picker In Your Life

Once you know when to use Fable, Mythos, Opus, Sonnet, and Haiku, the next hard choice is which Claude joke belongs on your shirt. CodeCulture's Claude-ing Is My New Addiction shirt is the casual pick for heavy Claude users, and the Professional Claude Whisperer shirt is the better fit for prompt engineers, AI coding leads, and anyone who has become the team's unofficial model translator.

Cost Comparison

The price jump is easy to miss because all the model names sound premium.

Model	Input price	Output price	Relative use case
Haiku 4.5	$1 / MTok	$5 / MTok	volume and speed
Sonnet 4.6	$3 / MTok	$15 / MTok	balanced daily work
Opus 4.8	$5 / MTok	$25 / MTok	complex reasoning
Fable 5	$10 / MTok	$50 / MTok	hardest general work
Mythos 5	$10 / MTok	$50 / MTok	trusted-access work

If a workflow produces long outputs, output pricing dominates. A 128k-output Fable run can be valuable, but it should be attached to a task where that much output has a real use.

FAQ

Is Fable 5 better than Opus 4.8?

Anthropic positions Fable 5 above the Opus class and describes it as the company's most capable widely released model. Opus 4.8 remains useful for complex tasks where Fable's extra cost, safeguards, or retention requirements are not worth it.

Is Mythos 5 available to everyone?

No. Mythos 5 is limited availability for approved Project Glasswing and trusted-access users. Fable 5 is the broadly available version of the same underlying model.

Which Claude model is best for coding?

Use Sonnet 4.6 for everyday coding, Opus 4.8 for hard coding and agentic reasoning, and Fable 5 for the hardest long-horizon coding tasks. Mythos 5 is not generally available and should not be treated as a normal coding upgrade.

Which Claude model is cheapest?

Claude Haiku 4.5 is the cheapest model in this comparison at $1 per million input tokens and $5 per million output tokens.

Which Claude model has the longest context?

Fable 5, Mythos 5, Opus 4.8, and Sonnet 4.6 are listed with 1M token context windows in Anthropic's model overview, with provider-specific caveats. Haiku 4.5 is listed at 200k.

The Bottom Line

Fable 5 is the new model to watch, but it should not replace every Claude workflow. Use Haiku for scale, Sonnet for the daily middle, Opus for complex work, Fable for the hardest generally allowed tasks, and Mythos only when your organization has approved trusted access.

The best teams will not ask "Which Claude model is smartest?" They will ask: "What is the lowest-cost model that can finish this job reliably, with the right policy fit?"

Sources

Claude Fable 5 vs Mythos 5: Same Model, Different Safeguards

Emcy — Wed, 10 Jun 2026 08:50:00 +0000

Originally published at https://codeculture.store/blogs/developer-culture/claude-fable-5-vs-mythos-5.

Claude Fable 5 vs Mythos 5: Same Model, Different Safeguards

Claude Fable 5 and Claude Mythos 5 are easy to misunderstand because the names sound like two different model families. They are not. Anthropic says they are the same underlying model. The difference is the access model and the safeguards around high-risk capabilities.

The simplest explanation: Fable 5 is the generally available version. Mythos 5 is the trusted-access version for approved cyberdefense, infrastructure, and eventually selected biology research users.

Key Takeaways

Fable 5 and Mythos 5 share the same underlying model.

Fable 5 is available broadly; Mythos 5 is restricted.

Fable 5 adds conservative classifiers for cyber, biology, chemistry, and distillation-related requests.

Mythos 5 lifts some safeguards for approved use cases through Project Glasswing and trusted access programs.

Most developers should use Fable 5 unless they are in an approved defensive security or research program.

Quick Comparison

Category	Claude Fable 5	Claude Mythos 5
Underlying model	Same as Mythos 5	Same as Fable 5
Availability	Generally available	Limited availability
Main audience	Developers, teams, knowledge workers, general Claude users	Approved cyberdefenders, infrastructure providers, selected researchers
API ID	`claude-fable-5`	`claude-mythos-5`
Context window	1M tokens	1M tokens
Max output	128k tokens	128k tokens
Thinking	Always-on adaptive thinking	Always-on adaptive thinking
Pricing	$10 input / $50 output per million tokens	$10 input / $50 output per million tokens
Safeguards	Conservative classifiers and routing	Some safeguards lifted for approved areas
Access path	Claude API, Claude Platform on AWS, Bedrock, Vertex AI, Microsoft Foundry	Project Glasswing and trusted access only

Sources: Anthropic announcement, Claude models overview, and Claude Platform release notes.

The Real Difference Is Policy, Not Intelligence

Fable 5 is not the "weaker" model in the normal sense. Anthropic says that Fable 5 and Mythos 5 are the same underlying model. For sessions where safeguards do not trigger, Anthropic says Fable 5's performance is effectively the same as Mythos 5.

The practical difference is what happens when a request enters sensitive territory.

Fable 5 uses classifiers around:

cybersecurity;
biology and chemistry;
distillation or reverse-engineering-style model extraction;
some reasoning extraction categories in API stop details.

When those systems trigger, the request may be routed, refused, or handled differently depending on the product surface. Anthropic says these safeguards are deliberately conservative and may catch some harmless requests. Source: Anthropic announcement.

Mythos 5 exists for users who need some of those restricted capabilities for legitimate work, especially defensive security and critical infrastructure. That is why Mythos 5 is tied to Project Glasswing and trusted access.

Why Anthropic Split The Names

The name split gives Anthropic a way to release one frontier capability tier into two policy environments.

For general users, the model is Fable. It can handle the hardest normal knowledge work and coding tasks, but it has safeguards around high-risk domains.

For approved users, the model is Mythos. It is aimed at tasks where the restricted capability is the point, such as advanced cyberdefense or approved life-science research.

That makes the product architecture unusually explicit:

Need	Model
Build an agent that can refactor code across a huge repo	Fable 5
Analyze complex documents, charts, tables, and product data	Fable 5
Run general Claude workflows with maximum available capability	Fable 5
Conduct approved advanced cyberdefense research	Mythos 5
Work inside Project Glasswing	Mythos 5
Use biology or chemistry capabilities under a trusted research program	Mythos 5 or trusted-access Fable configuration, depending on program

Which One Can Developers Actually Use?

Most developers can use Fable 5. Anthropic says Fable 5 is generally available beginning June 9, 2026 across the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry.

Mythos 5 is not self-serve. Anthropic's docs say it is limited availability for approved Project Glasswing customers. If a company needs access, it should contact Anthropic, AWS, or Google Cloud account teams. Source: Claude models overview.

For a normal engineering team, the decision is not really "Fable or Mythos?" It is "Should we pay for Fable, or should we stay on Opus, Sonnet, or Haiku for this workload?"

What Fable 5 Is Best For

Use Fable 5 when you want maximum general capability and can accept the safety policies.

Fable 5 fits:

large codebase analysis;
migration planning;
long-horizon Claude Code tasks;
complex spreadsheet or finance reasoning;
deep research synthesis;
vision-heavy app reconstruction;
multimodal product QA;
long-context knowledge work.

In CodeCulture terms, Fable 5 is a strong candidate for the "hard task model" in an AI coding command center. Use it when Sonnet or Opus can do the job eventually, but the work is expensive because the model loses track, needs too many turns, or fails to self-correct.

What Mythos 5 Is Best For

Use Mythos 5 only if you are an approved user with a legitimate need for capabilities that Fable 5 restricts.

Anthropic frames Mythos 5 around:

cyberdefense;
critical software infrastructure;
Project Glasswing;
trusted access for selected biology researchers;
high-end research and security workflows.

This matters because Mythos 5 is not a consumer upgrade path. It is closer to a controlled capability program. For most readers, it is useful to understand Mythos because it explains why Fable behaves the way it does.

Pricing Does Not Decide This Comparison

Fable 5 and Mythos 5 have the same listed price: $10 per million input tokens and $50 per million output tokens.

That means the difference is not cost. It is access and allowed use.

The more relevant cost comparison is Fable versus the lower Claude tiers:

Model	Input price	Output price
Fable 5	$10 / MTok	$50 / MTok
Opus 4.8	$5 / MTok	$25 / MTok
Sonnet 4.6	$3 / MTok	$15 / MTok
Haiku 4.5	$1 / MTok	$5 / MTok

Source: Claude models overview.

If you are not doing top-tier work, Fable is expensive. If it saves hours of senior engineering or research time, it may be cheap.

The Enterprise Catch: Retention

Anthropic's API release notes say Fable 5 requires 30-day data retention and is not available under zero data retention. That will matter for companies with strict data policies.

For a team evaluating Fable 5, ask:

Are we allowed to send this data to a model with 30-day retention?
Does our vendor policy allow this model class?
Are we working in a domain where classifiers may trigger often?
Do we need an audit trail for fallback, refusal, or routing behavior?
Can we route only high-value tasks to Fable and keep routine work on Sonnet or Haiku?

The model is powerful, but the policy surface is part of the product.

Wear The Claude Phase

If your team is now arguing about Fable, Mythos, Opus, and fallbacks in every planning meeting, you may as well dress for the era. The Claude-ing Is My New Addiction shirt fits the daily Claude power user, while the Professional Claude Whisperer shirt is for the person everyone asks to write the prompt when the first five attempts fail.

Fable 5 vs Mythos 5 FAQ

Is Mythos 5 more powerful than Fable 5?

Not in the usual model-quality sense. Anthropic says Fable 5 and Mythos 5 are the same underlying model. Mythos 5 differs because some safeguards are lifted for approved users and approved use cases.

Can I sign up for Mythos 5?

Not through normal self-serve access. Anthropic says Mythos 5 is available to approved Project Glasswing customers and planned trusted access programs. Most developers should use Fable 5.

Why does Fable 5 sometimes fall back or refuse?

Fable 5 uses classifiers for high-risk categories, including cyber, biology, chemistry, and distillation. Depending on the surface, a flagged request may be routed, refused, or handled by another model.

Do Fable 5 and Mythos 5 cost the same?

Yes. Anthropic lists both at $10 per million input tokens and $50 per million output tokens.

Which should I write about for SEO?

Write about both, but make Fable 5 the primary keyword. Fable is the model most readers can use. Mythos is the comparison term that explains the release strategy.

The Bottom Line

Claude Fable 5 is the broadly available Mythos-class model. Claude Mythos 5 is the controlled-access version for approved high-risk and high-value work.

The useful framing is simple: Fable is capability plus safeguards. Mythos is capability plus trusted access. For most developers and teams, the action item is not to chase Mythos. It is to test whether Fable 5 can handle the long, messy work that previous models could not finish reliably.

Sources

Claude Fable 5 Explained: What Anthropic's New Mythos-Class Model Means

Emcy — Wed, 10 Jun 2026 08:48:16 +0000

Originally published at https://codeculture.store/blogs/developer-culture/claude-fable-5-explained.

Claude Fable 5 Explained: What Anthropic's New Mythos-Class Model Means

Claude Fable 5 is Anthropic's new top-tier general-access Claude model, launched on June 9, 2026. The short version: Fable 5 brings Mythos-class capability to regular Claude users, but with safeguards that route some sensitive cyber, biology, chemistry, and distillation-related requests away from Fable and toward Claude Opus 4.8 or refusal handling, depending on the surface.

That makes Fable 5 more than another model bump. It is Anthropic's first broad attempt to release a model above the Opus tier while still controlling the parts of the model that create the highest misuse risk.

Key Takeaways

Claude Fable 5 is Anthropic's most capable widely released model as of June 10, 2026.

It shares an underlying model with Claude Mythos 5, but Fable adds stricter safeguards for general use.

Developers get a 1M token context window, 128k max output, always-on adaptive thinking, and the API ID claude-fable-5.

The main tradeoff is safety routing: some harmless advanced requests may be caught by classifiers.

For most developers, Fable 5 is the model to test on long-horizon coding, research, analysis, and agent workflows.

What Is Claude Fable 5?

Claude Fable 5 is a Mythos-class Claude model that Anthropic has made generally available. Anthropic describes Mythos-class models as sitting above the Opus class in capability. Fable 5 is the general-release version of this capability tier.

In Anthropic's launch post, the company says Fable 5 performs especially well on long and complex tasks, including software engineering, knowledge work, vision, scientific research, and autonomous agent-style workflows. Source: Anthropic announcement.

For developers, the concrete API details matter:

Feature	Claude Fable 5
API model ID	`claude-fable-5`
Availability	Generally available beginning June 9, 2026
Context window	1M tokens
Max output	128k tokens
Thinking mode	Always-on adaptive thinking
Pricing	$10 per million input tokens, $50 per million output tokens
Main caveat	Classifier safeguards on some sensitive request categories

Source: Claude models overview and Claude Platform release notes.

Why Fable 5 Matters

The biggest shift is not only raw intelligence. It is task duration.

Most AI model launches promise better reasoning, better coding, and better instruction following. Fable 5 is positioned around longer-horizon work: tasks where the model has to plan, preserve context, use tools, recover from mistakes, and continue working after the obvious first answer is exhausted.

That matters for developers because the hardest AI coding work is rarely "write this function." It is:

migrate a large codebase without breaking conventions;
reason across hundreds of files;
compare screenshots with implementation details;
keep a multi-step plan coherent over time;
decide when to test, refactor, or stop;
produce documentation that actually matches the diff.

Anthropic highlights early customer testing from teams using Fable 5 for codebase migrations, complex analytics, legal redlines, finance reasoning, coding benchmarks, and long-running production workflows. Treat those examples as vendor-provided signals, not neutral benchmark proof, but they point at the model's intended job: high-autonomy work where smaller models lose the thread.

The Safety Story Is The Product Story

Fable 5 exists because Mythos-class models are powerful enough to need extra controls. Anthropic says the model's capabilities in cybersecurity, biology, chemistry, and distillation create misuse risk. The company therefore routes certain requests away from Fable 5.

In the launch post, Anthropic says the safeguards trigger in less than 5% of sessions on average. When they do trigger in some Claude surfaces, a request may be handled by Claude Opus 4.8 instead. In the API, Anthropic documents classifier refusals through stop_reason: "refusal" and related stop details. Source: Claude Platform release notes.

This is the tension developers need to understand:

Benefit	Tradeoff
Access to Anthropic's strongest broadly available model	More conservative safety filters
Better long-context work	Higher token cost than Opus, Sonnet, or Haiku
Stronger agentic coding and analysis	Some advanced research tasks may route away or refuse
Broad cloud availability	30-day retention requirement for Fable traffic

The 30-day retention requirement is especially important for enterprise teams. Anthropic states that Fable 5 is not available under zero data retention on the Claude API. Source: Claude Platform release notes.

Who Should Use Claude Fable 5?

Use Fable 5 when the work is hard enough to justify the cost.

Good Fable 5 use cases:

large refactors and codebase migrations;
long-running agent tasks in Claude Code or custom tooling;
complex product analysis across many documents;
research synthesis where the model must hold a long chain of evidence;
vision tasks that require precise interpretation;
multi-file debugging where cheap models keep missing the causal link;
executive analysis where the output has to combine charts, tables, policy, and reasoning.

Fable 5 is probably overkill for:

simple summarization;
everyday copy edits;
small code snippets;
support macros;
short classification tasks;
low-risk autocomplete.

For those jobs, Sonnet or Haiku may still be the better default because they are cheaper and faster.

What Developers Should Test First

If you want to ride the Fable 5 trend without writing generic launch coverage, test it on the kind of work developers actually care about.

Start with three evaluations:

A long refactor: ask Fable 5 to inspect an existing codebase, identify the migration plan, make a scoped patch, and explain the tradeoffs.
A screenshot-to-code task: give it a product screenshot and ask for a component implementation that matches the UI system.
A research-to-implementation task: give it docs, a bug report, and source files, then ask it to produce a tested fix.

The goal is not to prove that Fable 5 is "best." The goal is to learn whether the model stays coherent after the first few steps. That is where frontier models separate themselves in real workflows.

Claude Gear For The Fable 5 Era

If Fable 5 turns Claude from "useful assistant" into "daily work habit" for you, CodeCulture has two shirts that fit the moment: Claude-ing Is My New Addiction for the builders who keep one more Claude tab open, and Professional Claude Whisperer for the teammate who somehow gets the model to understand every vague product request.

Claude Fable 5 FAQ

Is Claude Fable 5 real?

Yes. Anthropic announced Claude Fable 5 and Claude Mythos 5 on June 9, 2026. Fable 5 is generally available, while Mythos 5 is restricted to Project Glasswing and trusted access programs.

Is Fable 5 better than Opus?

Anthropic positions Fable 5 above the Opus class. The practical answer is that Fable 5 should be tested first for the hardest long-context, agentic, coding, research, and vision work. Opus 4.8 may still be a better fit when you want fewer safeguards, lower cost, or established Opus-tier behavior.

What is the Fable 5 API model ID?

The Claude API model ID is claude-fable-5.

Does Fable 5 have a 1M token context window?

Yes. Anthropic's model overview and release notes list a 1M token context window for Claude Fable 5 and Claude Mythos 5.

What is the biggest downside of Fable 5?

The biggest downside is the combination of higher cost, new safety routing, and a 30-day data retention requirement. For sensitive enterprise workflows, that retention policy may matter as much as model quality.

The Bottom Line

Claude Fable 5 is Anthropic's new flagship for general users. The interesting part is not just that it is stronger. It is that Anthropic is trying to make Mythos-class capability broadly usable without releasing every capability in every context.

For developers, the right move is simple: test Fable 5 on work where long context, autonomy, and recovery from mistakes actually matter. If it only saves a few seconds on a small prompt, use a cheaper model. If it keeps a hard task coherent for hours, that is the story.

Sources

I replaced a SaaS tool with a Python script because my ad spend was embarrassing

Emcy — Fri, 15 May 2026 15:38:00 +0000

Last year I opened my Meta Ads dashboard, did the math, and just stared at the screen for a second.

I had spent more on running ads that week than the store had made.

Not close. Actually more.

So yeah. That's where this series starts.

A bit of context

I run Code Culture on the side. It's a developer apparel brand, shirts with terminal jokes on them, hoodies for the kind of person who has opinions about tab width. I work a full-time job in data during the day and do this in whatever's left over.

Running it solo means every subscription you're paying for is money that could've been product, or ads, or just staying in your pocket. And for a while I was paying for a SaaS tool that bulk uploads ad creatives to Meta. It was fine. It did the job. But it was also a monthly cost on top of an ad budget that clearly wasn't working yet.

At some point I just stopped and thought about what the tool was actually doing. And the answer was: it uploads your images to Meta, creates a creative, wires it to an ad set, saves it as a draft. That's the whole thing.

So I built it myself.

What I made

I wrote meta_ad_uploader.py with Claude over two evenings. It talks to the Meta Graph API and does exactly what I needed:

Upload the image to the ad account's image library
Find or create the ad set
Build the creative with copy, headline, CTA, and UTM params
Create the ad as paused. Nothing goes live on its own. That last bit was the non-negotiable part. I didn't want a script that could accidentally start spending money at 2am because I forgot to double-check something. Every ad it creates is a draft. I go in manually, look at it, then activate.

The thing I'm weirdly proud of:

I was testing a bunch of different creative angles at the same time. Lifestyle photos, product shots, stuff leaning into specific dev jokes. Each one needed different primary copy to match the vibe of the image.

I didn't want to build a whole config system for this. I just wanted something simple that worked. So I went with filename matching.

AD_COPY = {
    "chaos": (
        "Some developers have a staging environment.\n\n"
        "Some have a rollback plan.\n\n"
        "Some have a very good story about why production was down for 47 minutes on a Thursday.\n\n"
        "This shirt is for the last group."
    ),
    "conference": (
        "There's a specific moment at every tech conference.\n\n"
        "Someone walks in and another developer across the room sees their shirt and immediately understands.\n\n"
        "Not everyone gets it. That's the point."
    ),
    # ... more angles
    "default": "Premium developer apparel. Built for the ones who stay calm when the monitors go red.",
}

def match_copy(filename: str) -> str:
    name = Path(filename).stem.lower()
    for keyword, copy in AD_COPY.items():
        if keyword != "default" and keyword in name:
            return copy
    return AD_COPY["default"]

Name your file chaos_v3.jpg and it picks up the "testing in prod" copy. Name it conference_final.png and it gets the conference angle. Rename your files, change your copy. No database, no UI, no config file.

There's also a --list-copy flag that prints which copy each file would get before you run anything. Genuinely useful when you've got 10 images and you want to sanity-check it first.

One thing the Meta docs don't make obvious

When you upload an image through the API, you get back a hash, not an ID. And that hash is what you need when you build the creative. I spent more time than I'd like to admit figuring that out because the docs kind of gloss over it.

def upload_image(image_path: Path) -> str:
    mime_type, _ = mimetypes.guess_type(str(image_path))
    with open(image_path, "rb") as f:
        resp = requests.post(
            f"{BASE_URL}/act_{AD_ACCOUNT}/adimages",
            data={"access_token": META_TOKEN},
            files={"filename": (image_path.name, f, mime_type)},
        )
    resp.raise_for_status()
    data = resp.json()
    images = data.get("images", {})
    entry = next(iter(images.values()), None)
    if not entry:
        raise RuntimeError(f"Image upload failed: {data}")
    return entry["hash"]  # this is what you need, not an "id"

Build the dry-run first, not last

I added --dry-run before I got the rest of the script working, and it saved me from creating a bunch of garbage draft ads while I was still figuring out the API shape.

# see which copy each image gets, no API calls
python3 meta_ad_uploader.py --list-copy chaos_v1.jpg conference_v2.jpg

# print the full payload without posting anything
python3 meta_ad_uploader.py --dry-run chaos_v1.jpg

# actually run it
python3 meta_ad_uploader.py --folder /path/to/creatives/

If you're building anything that touches a paid API or creates real objects in someone else's system, just build the dry-run path first. You'll thank yourself.

Did it work

The script works fine. I drop images into a folder, run it, and a couple minutes later I've got paused draft ads sitting in Ads Manager ready to review.

The SaaS subscription is gone.

The part that actually surprised me wasn't the code. The Meta API is reasonably documented and Claude handled most of the boilerplate. What surprised me was how much headspace the SaaS tool had been using. You don't fully trust something you don't understand, so you end up double-checking things, second-guessing the output, wondering if it did what you think it did. Building it yourself removes that whole layer.

I still spent too much on ads that didn't convert. That's a targeting and creative problem and no script fixes it. But at least when something goes wrong now, I know exactly where to look.

This is the first post in a series I'm writing about building Code Culture with AI tools. The store is developer apparel. The operational stack is Python, Claude, and a lot of trial and error. The budget is not VC-backed.

More coming.