Claude Opus 5 for Marketing: What’s New and How to Use It

Anthropic released Claude Opus 5 on July 24, 2026 at $5 per million input tokens and $25 per million output tokens, the same list price as the Opus 4.8 it replaces. If you run marketing, that flat price is the story. Better ad copy is not: every capable model has written competent ad copy for a while now, and treating that as the upgrade is the fastest way to overpay for work a cheaper model already does well.

Three things change your operating math. You can set a cost and capability dial per request, which buys flagship reasoning for the hero asset and cheap thinking for the hundredth ad variant, out of one model and one integration. The model checks its own work before it hands it back, so a pipeline runs further before a person has to look at it. And a one million token window makes grounding in your own brand material practical rather than aspirational. Everything worth planning around comes from those three.

Quick answers:

• Is Claude Opus 5 real, and when was it released?
• How much does Claude Opus 5 cost?
• What is the Claude Opus 5 context window?
• What changed between Claude Opus 4.8 and Claude Opus 5?
• Should a marketing team use Claude Opus 5 or Fable 5?
• Can Claude Opus 5 generate images or ad creative?
• What is the Claude Opus 5 effort parameter, and why does it matter for advertising?
• Where can I access Claude Opus 5?

What Anthropic actually shipped on July 24, 2026

Claude Opus 5 is real, it is out, and the model identifier is claude-opus-5. Anthropic announced it and made it available the same day across the Claude API, Claude.ai, Claude Code, Claude Cowork, Amazon Bedrock, the Claude Platform on AWS, Google Cloud, and Microsoft Foundry. On the consumer plans it became the default model on Claude Max and the strongest model available on Claude Pro.

Anthropic’s own framing is a step-change improvement for the Opus tier: stronger coding, more capable agents, sharper professional work. Its model documentation goes further and names Opus 5 the recommended place to start for complex agentic coding and enterprise work, reserving Fable 5 for cases where you need the highest capability available at any price. That detail is easy to skim past and worth sitting with. The company is telling you that its second-most-capable model is the one most teams should default to.

The specifications Anthropic publishes are straightforward. One million tokens of context, which Anthropic puts at roughly 555,000 words. Maximum output of 128,000 tokens, and up to 300,000 through a Batch API beta. A reliable knowledge cutoff of May 2026. Text and image input, text output. Latency that Anthropic’s own comparison table describes as moderate, sitting between the faster Sonnet 5 and the slower Fable 5.

What is genuinely new versus Opus 4.8

Four changes, and only one of them is about raw intelligence.

The effort parameter is the one that will change how you work. It takes three values, low, medium, and high, and it lets you decide per call how much thinking the model spends before answering. It defaults to high on the Claude API and in Claude Code, which means if you never touch it you are paying for maximum effort on every request, including the trivial ones. Fortune, covering the launch, treated this cost and capability toggle as the headline feature of the release. From an advertising chair that reading is correct, and we will come back to why.

Fast mode is the second lever. Anthropic says it runs at roughly 2.5 times the default speed for roughly twice the base price. Note the shape of that trade: you are buying latency with money, not capability with money. Those are different purchases and they suit different jobs.

Agentic self-verification is the third. Anthropic states that Opus 5 verifies its own work and recovers from errors without the user stepping in, which cuts the back-and-forth a multi-step task needs. TechCrunch’s launch-day report quotes the announcement describing the model as much stronger at verifying its work and iterating carefully until it succeeds. This is a claim about workflow, not about prose quality, and it is the claim most likely to matter to anyone building a pipeline rather than a chat habit.

Fourth, performance. On Anthropic’s own published benchmarks the gains are large. It reports that Opus 5 more than doubles Opus 4.8’s score on Frontier-Bench at a lower cost per task, lands within half a percent of Fable 5’s peak CursorBench score at half the cost, scores three times as high as the next-best model on ARC-AGI 3, passes roughly one and a half times the rate of the next-best model on Zapier’s AutomationBench, and beats Fable 5’s best OSWorld 2.0 result at just over a third of the cost. Every one of those figures is Anthropic-published, on benchmarks Anthropic selected. Hold them the way you would hold your own case-study numbers: a direction of travel, not a measurement you can build a budget on.

The pricing math that changed, and the part that did not

The list price is $5 per million input tokens and $25 per million output tokens. That is exactly what Opus 4.8 cost. The price did not move.

This is the whole reason the launch is interesting rather than merely notable. Normally a capability jump arrives with a price jump, and the decision in front of you is whether the new tier is worth its premium. Here there is no premium to weigh. If your team already budgets for Opus, the migration is a change to the model string, and your cost model is unchanged. Anthropic’s documentation now lists Opus 4.8 as the prior generation and points users at Opus 5, so there is not much of a case for staying.

Three discounts move the effective rate, and marketing workloads happen to be unusually good at using all three. Prompt caching can cut costs by up to 90 percent according to Anthropic, and it applies exactly where campaign work repeats itself: the same brand guidelines, the same product catalogue, the same compliance rules attached to every call. Batch processing runs at a 50 percent discount and suits overnight variant generation and bulk analysis, where nobody is waiting on the answer. US-only inference is billed at 1.1 times list, which matters if your data residency rules require it.

Set against Fable 5 at $10 in and $50 out, Opus 5 is roughly half the price of the top tier. Anthropic’s published CursorBench result puts it within half a percent of Fable 5’s peak. You do not have to accept that number at face value to notice what the pricing implies about the company’s own view of where most work belongs.

The strategy behind that pricing is not a secret. CNBC’s launch coverage quotes Dianne Penn, Anthropic’s head of product management for research, saying enterprises are looking for value, and that a cheaper model which is not accomplishing a similar level of quality is actually not useful. Read that as a statement about the market, not about the model. Buyers have stopped paying premium rates to experiment, and the vendors have noticed. That is why a capability jump arrived at a flat price, and it is good news for anyone whose AI spend has to justify itself against a cost per lead.

Where Opus 5 sits in the Claude lineup

Where Opus 5 Sits in the Claude Lineup

Six models, three price tiers. All figures are Anthropic’s own published prices and specs as of July 24, 2026.

Top tier

Fable 5

• Model id: claude-fable-5
• Price: $10 in / $50 out per million tokens
• Context: 1M tokens. Latency: slower.
• Anthropic’s most capable widely released model.
• Reach for it for long-running agents and the hardest calls.

Highest capability, highest price

Top tier, restricted

Mythos 5

• Model id: claude-mythos-5
• Price: $10 in / $50 out per million tokens
• Context: 1M tokens.
• Same specs as Fable 5, but invite-only.
• Aimed at defensive cybersecurity work (Project Glasswing).

Not a marketing tool

The new default

Opus 5

• Model id: claude-opus-5
• Price: $5 in / $25 out per million tokens
• Context: 1M tokens. Latency: moderate.
• Announced and available July 24, 2026.
• Anthropic’s recommended start for complex agentic and enterprise work.

Same price as Opus 4.8

Workhorse

Sonnet 5

• Model id: claude-sonnet-5
• Price: $3 in / $15 out per million tokens
• Introductory rate of $2 in / $10 out through August 31, 2026.
• Context: 1M tokens. Latency: fast.
• Positioned as the best balance of speed and intelligence.

Volume work

Cheapest

Haiku 4.5

• Model id: claude-haiku-4-5
• Price: $1 in / $5 out per million tokens
• Context: 200k tokens. The smallest window in the lineup.
• The fastest model Anthropic ships, described as near-frontier.
• For high-volume, low-judgment tasks.

Fastest, smallest window

Prior generation

Opus 4.8

• Model id: claude-opus-4-8
• Price: $5 in / $25 out per million tokens
• Context: 1M tokens. Latency: moderate.
• Identical price to Opus 5, now the older model.
• Anthropic’s docs direct users to migrate to Opus 5.

Legacy

The official steer is short: start with Opus 5 for complex agentic and enterprise work, and reach for Fable 5 when you need the highest available capability. Underneath that, the lineup sorts cleanly by job. Haiku 4.5 at $1 in and $5 out is for high volume and low judgment, and it is the one model here with a smaller 200,000 token window. Sonnet 5 at $3 in and $15 out, currently discounted to $2 and $10 through August 31, 2026, is the balanced workhorse. Opus 5 is the reasoning tier you can now afford to leave running. Fable 5 and the invite-only Mythos 5 sit above it at double the price, and Mythos 5 is aimed at defensive cybersecurity rather than at anything a marketing team does.

When Fable 5 launched we argued that the discipline was to give your most expensive model only the hard decisions and push everything else down the stack. That argument still holds, but Opus 5 changes where the line falls. Tiering used to mean running two or three models and maintaining the routing between them. The effort parameter moves a large part of that tiering inside a single model, which is a simpler thing to operate and a much easier thing to reason about when something breaks.

Seven ways marketers and agencies can actually use it

Six Marketing Uses, and the Feature Behind Each One

If a use case does not trace back to a named Opus 5 capability, it is a generic language-model use and any capable model will do it.

Feature: agentic self-verification

Run a content or ad pipeline end to end

Anthropic says Opus 5 checks its own work and recovers from errors without the user stepping in. That is what lets a brief to research to draft to quality-assurance chain run with fewer human handoffs, rather than stopping for a person at every stage.

Fewer handoffs, not better prose

Feature: the effort parameter and fast mode

Flagship quality here, cheap fan-out there

Effort accepts low, medium, or high, and defaults to high on the API and in Claude Code. Fast mode runs at roughly 2.5 times the default speed for roughly twice the base price. One model, several unit costs: full effort on the hero page, low effort on the hundredth ad variant.

The lever that matters most

Feature: 1M-token context

Briefs grounded in your own material

A one million token window holds a full brand guideline set, past campaign exports, competitor pages, and a persona library in one conversation. The brief comes back grounded in your documents instead of in the model’s general training.

Grounding, not recall

Feature: image input (vision)

Critique competitor creative and pages

Opus 5 accepts images as input, so you can hand it competitor ad creatives or landing-page screenshots and get a read on layout, messaging, and hierarchy. This is analysis of images. It is not generation of them.

Reads images, does not make them

Feature: coding and artifact outputs

Working landing-page prototypes

Anthropic’s headline claim for this release is stronger coding, and it describes much stronger visual outputs, meaning generated artifacts, diagrams, and interface code. That makes it a better producer of working HTML prototypes, and still not an image generator.

Code and artifacts, not photos

Feature: reasoning plus large context

Read your own performance data

Paste raw advertising or Search Console exports into a window big enough to hold them, and the gains Anthropic claims on professional reasoning go toward pattern analysis and next-test hypotheses rather than toward another draft of a headline.

Hypotheses, not dashboards

Six of these trace to a named capability. The seventh is the honest one.

One, run an agentic content or advertising pipeline end to end. The self-verification claim is what makes a brief to research to draft to quality-assurance chain viable with fewer human checkpoints. This is the use case with the most upside and the most room to fool yourself, and we come back to it below.

Two, split effort across a campaign build. Full effort on the strategy and the hero landing page, low effort or fast mode on the long tail of variants and metadata. One model, several unit costs.

Three, build briefs grounded in your own material. A one million token window will hold your brand guidelines, a year of campaign exports, a competitor set, and your persona library at once. The output comes back built from your documents rather than from the model’s general sense of your category.

Four, critique creative and landing pages. Image input means you can hand it competitor ads or page screenshots and get a read on layout, messaging, and hierarchy. This is the fastest honest win in the list, because reviewing creative is a job most teams do inconsistently and it needs no new tooling.

Five, produce working prototypes. Stronger coding and artifact generation make it a better producer of functioning HTML landing-page prototypes, which is a workflow we have written about at length in our guide to building lead-gen landing pages with a page contract.

Six, read your own performance data. Drop raw advertising or Search Console exports into a window large enough to hold them and ask for patterns and next-test hypotheses rather than for a summary. This pairs well with the diagnostic work in our Google Ads for lead generation guide and our breakdown of Google Ads bid strategies. It earns its keep hardest during a platform migration, when you have to re-plan an account against rules that just changed, which is exactly the situation created by Local Services Ads moving into Google Ads.

Seven, everything else: copy, headlines, email sequences, social captions. These work well. They also worked well on the last model, and they work well on cheaper models today. Put them on the list, price them accordingly, and do not let them carry the business case for an upgrade.

The effort dial in practice: the economics of a variant fan-out

Here is where the theory meets an actual campaign build, because a campaign is not one task. It is one hard task and several hundred easy ones.

Writing the positioning, choosing the offer, and deciding what the hero page argues are jobs where being slightly better is worth real money, and where you want every bit of reasoning you can buy. Filling out a responsive search ad is not that job. Google’s own documentation lets you supply up to 15 headlines and 4 descriptions per responsive search ad, and says the more you provide, the more combinations it can test. Multiply that across ad groups and you are producing hundreds of assets that need to be competent, on-brand, and compliant, and past that the marginal value of extra thinking is close to zero. If you have built responsive search ads with pinning and dynamic keyword insertion, you already know the shape of this: a small number of decisions that matter enormously, wrapped in a large amount of production that has to be correct rather than brilliant.

Before this release, you handled that split in one of two unsatisfying ways. Pay flagship rates for the routine production, which works and quietly wastes money. Or route the routine work to a cheaper model, which saves money and buys you a second integration, a second prompt set, and a second set of failure modes to debug at eleven at night.

The effort parameter collapses that choice. Same model, same prompts, same integration, different amounts of thinking per call. High effort on the brief and the hero asset. Low effort on the fan-out. Fast mode where a human is waiting on the answer and you are willing to pay double the base rate for roughly two and a half times the speed, which is a genuinely reasonable trade during a live campaign build and a poor one for an overnight batch job that nobody is watching.

The trap is the default. Effort defaults to high on the API and in Claude Code, so a pipeline that never sets it is buying maximum reasoning for every metadata field it generates. At volume that is not a small leak. Auditing which of your calls actually need high effort is an afternoon of work, and it is the highest-return thing you can do with this release.

What running an agentic pipeline actually teaches you

Elevarus runs its own content and advertising operation as an agentic system on Claude models. Not as a demo, as the thing that does the work. That gives us a view of the self-verification claim that a benchmark does not, and it cuts both ways.

Self-verification is real, and it removes a genuine class of failure: the model catching its own broken link, its own malformed schema, its own contradiction between paragraph two and paragraph nine. That is the tedious checking that used to force a human handoff at every stage, and removing it is why a pipeline now runs further before someone has to look at it. Anthropic’s framing of fewer round trips matches what the work feels like.

The part nobody puts in a launch post is that a model checking its own work is still one party checking itself. It is very good at catching mechanical faults and structurally poor at catching the thing it was confidently wrong about, because the same judgment produced the error and the review. So the checks that matter most in our pipeline are the ones the model does not perform on itself: an independent pass, run separately, against a standard the drafting step cannot see. We wrote up that architecture in our piece on building a content quality-control pipeline, and the principle survived this upgrade unchanged.

One practical note for anyone running pipelines in a regulated vertical. If you advertise health insurance, final expense, or anything else where the copy brushes up against sensitive subject matter, you will occasionally trip a model’s safety classifier, and historically that returned an error and broke the run. TechCrunch reports that Anthropic is rolling out an opt-in beta called Automatic Fallbacks, which routes a request that trips a classifier to a less powerful model so the caller gets a usable response instead of a failure. That is a small feature with an outsized effect on whether an unattended pipeline survives the night.

If that posture sounds familiar, it is the same one we take toward lead quality. We do not accept a lead’s own account of itself, we verify it independently before it costs anybody money. A self-verifying model earns the same treatment. Better self-checking means your independent gates catch less noise and can spend their attention on the failures that actually matter. It does not mean you remove them. Anyone selling you an autonomous pipeline with no independent verification layer is selling you the thing we spend our days telling lead buyers not to buy.

Vision means it reads your creative, not makes it

One clarification to plan around. Opus 5 takes images as input, so you can hand it a competitor’s ad, a landing-page screenshot, or your own creative and get back a read on layout, messaging, and hierarchy, or the copy pulled straight out of it. What it produces is text, code, and artifacts, not photographs or finished ad images. Anthropic’s phrase “much stronger visual outputs” means generated artifacts, diagrams, charts, and interface code. So display banners and social creative still need a separate image tool and a separate budget line, and Opus 5 is the thing that tells you what is wrong with the creative you already have.

How to get started

Through the API, use the model identifier claude-opus-5. It is a dateless pinned snapshot, so there is no date suffix to track. On Amazon Bedrock the identifier is anthropic.claude-opus-5. It is also available through the Claude Platform on AWS, Google Cloud, and Microsoft Foundry, and 9to5Mac’s launch-day coverage confirms the same same-day availability across the consumer apps.

If you want to use it without writing code, it is in Claude.ai, where it is the default on Claude Max and the strongest option on Claude Pro, and in Claude Code and Claude Cowork.

Three things to do in your first week. Set the effort parameter explicitly everywhere, rather than inheriting the high default. Turn on prompt caching for the context you resend on every call, which for most marketing teams is brand guidelines and compliance rules. And move anything that is not latency-sensitive to the batch API at half price. None of those three require you to change what your pipeline does. They change what it costs, which is usually the constraint that decides whether an idea gets built at all.

The one thing to take from this

The story of this release is not that the model got smarter, though Anthropic’s numbers say it did. It is that the price of the reasoning tier held flat while the control over how much reasoning you buy got much finer. Set the effort dial deliberately, cache the context you resend on every call, and batch the work nobody is waiting on. Do those three and an agentic marketing system stops being a pilot you fund out of curiosity and becomes something you can afford to leave running. Running all the time is where these systems stop being interesting and start being useful.

Most funnels do not need a smarter model. They need better lead generation economics, and the hard part is telling the two apart. Book a free call and we will walk your campaign and content operation with you, then say plainly which one you are looking at.

Frequently Asked Questions

Is Claude Opus 5 real, and when was it released?

Yes. Anthropic announced Claude Opus 5 on July 24, 2026, and made it available the same day. The exact model identifier is claude-opus-5. It shipped simultaneously across the Claude API, Claude.ai, Claude Code, Claude Cowork, Amazon Bedrock, the Claude Platform on AWS, Google Cloud, and Microsoft Foundry. On the consumer plans it became the default model on Claude Max and the strongest model available on Claude Pro. Anthropic’s own one-line positioning calls it a step-change improvement for the Opus tier, with stronger coding, more capable agents, and sharper professional work, and its documentation names Opus 5 the recommended place to start for complex agentic coding and enterprise work. If you saw earlier chatter about codenames or rollout dates before that day, disregard it. The official announcement date is July 24, 2026, and that is the only date worth building a plan around.

How much does Claude Opus 5 cost?

Five dollars per million input tokens and twenty five dollars per million output tokens. The number that matters for planning is that this is exactly the same list price as Opus 4.8, the model it replaces, so a team already budgeting for Opus does not need a new number in the spreadsheet. Three things move the effective rate. Fast mode, which Anthropic says runs at roughly two and a half times the default speed, costs roughly twice the base price. Prompt caching can cut costs by up to 90 percent according to Anthropic, which matters a great deal if you send the same brand guidelines or product catalogue on every call. Batch processing is priced at a 50 percent discount, which suits overnight variant generation and bulk analysis where you do not need an answer in the next second. US-only inference is billed at 1.1 times the list rate. For comparison, Fable 5 sits at ten dollars in and fifty dollars out, so Opus 5 is roughly half the price of Anthropic’s top tier.

What is the Claude Opus 5 context window?

One million tokens, which Anthropic describes as roughly 555,000 words or about 2.5 million characters. Maximum output is 128,000 tokens, and up to 300,000 tokens through a Batch API beta header. The reliable knowledge cutoff is May 2026. For a marketer the practical meaning of a one million token window is that grounding stops being a rationing exercise. You can put a complete brand guideline document, a year of campaign exports, a set of competitor landing pages, and your persona library into a single conversation and ask for a brief, and the answer is built from your material rather than from the model’s general sense of your category. That is a different mode of work from pasting in a paragraph of context and hoping. It is worth saying plainly that a bigger window does not by itself make the output better. It makes the output grounded, which is a different and more useful property.

What changed between Claude Opus 4.8 and Claude Opus 5?

Four things, and only one of them is about raw intelligence. First, the effort parameter, which accepts low, medium, or high and lets you trade cost against capability on a per-call basis, defaulting to high on the Claude API and in Claude Code. Second, fast mode, which Anthropic says runs at roughly two and a half times the default speed for roughly twice the base price. Third, the agentic behaviour: Anthropic states that Opus 5 verifies its own work and recovers from errors without the user intervening, which reduces the back-and-forth a multi-step workflow needs. Fourth, performance. On Anthropic’s own published benchmarks Opus 5 roughly doubles Opus 4.8’s results, in one case at a lower cost per task. Treat those benchmark figures as vendor-published, because they are, and because the benchmarks were selected by the company selling the model. The price did not change at all. Same five dollars in, same twenty five dollars out, same one million token context window.

Should a marketing team use Claude Opus 5 or Fable 5?

Start with Opus 5. That is Anthropic’s own steer, and for marketing work it is usually the right one. Its documentation says to start with Opus 5 for complex agentic coding and enterprise work, and to reach for Fable 5 when you need the highest available capability. The price gap is the whole argument: Fable 5 is ten dollars in and fifty dollars out, Opus 5 is five dollars in and twenty five dollars out, so you are paying double for the top tier. On Anthropic’s published CursorBench results, Opus 5 landed within half a percent of Fable 5’s peak score at half the cost. The practical pattern for a marketing team is to run Opus 5 as the working model for pipelines, briefs, analysis, and production, and keep Fable 5 for the small number of genuinely hard, one-off strategic calls where the cost of being wrong is much larger than the cost of the tokens. We wrote about that tiering discipline in more depth when Fable 5 launched.

Can Claude Opus 5 generate images or ad creative?

No, and this is the single most common misreading of the launch. Opus 5 accepts images as input, which is called vision, so you can upload a competitor’s ad, a screenshot of a landing page, or your own creative and have the model analyse layout, messaging, and visual hierarchy or pull the copy out of it. What it does not do is produce photographic images. Anthropic’s announcement refers to much stronger visual outputs, and that phrase means generated artifacts, diagrams, charts, and user interface code, not a finished ad image. If your plan for Opus 5 involves generating display banners or social creative, that plan needs a separate image-generation tool, and you should budget for it separately. Read the capability the right way round: Opus 5 is unusually good at looking at your creative and telling you what is wrong with it, and it cannot make you a new one.

What is the Claude Opus 5 effort parameter, and why does it matter for advertising?

The effort parameter accepts three values, low, medium, and high, and it lets you decide how much thinking the model spends on a given call. It defaults to high on the Claude API and in Claude Code, so if you never set it, you are paying for maximum effort on every request, including the trivial ones. Fortune, covering the launch, framed this cost and capability toggle as the headline feature of the release, and for advertising work that framing is right. The reason is fan-out. A campaign build is not one task, it is one hard task and several hundred easy ones: the strategy and the hero asset genuinely need full reasoning, while the ninetieth headline variant, the metadata, and the routine formatting do not. Before this control you either paid flagship rates for the routine work or moved it to a different, cheaper model and managed two integrations. The effort dial puts that tiering inside one model, which is a meaningfully simpler thing to operate.

Where can I access Claude Opus 5?

Everywhere Anthropic ships, and it went live on all of them on launch day. Through the Claude API using the model identifier claude-opus-5. In the Claude apps at Claude.ai, where it is the default model on Claude Max and the strongest option on Claude Pro. In Claude Code and in Claude Cowork. And through the major cloud platforms: Amazon Bedrock, where the identifier is anthropic.claude-opus-5, the Claude Platform on AWS, Google Cloud, and Microsoft Foundry. The identifier is a dateless pinned snapshot, so you do not need to track a date suffix. If your team already runs on Opus 4.8, the migration is usually a one-line change to the model string, and Anthropic’s documentation now lists Opus 4.8 as the prior generation and points users to Opus 5. Since the price is identical, that swap does not change your cost model at all.