Anthropic launched Claude Opus 5 at $5/$25: when to migrate from Sonnet 5 (and when not)

August 2, 202615 minAnthropic, Claude, Opus 5, Sonnet 5, AI model pricing, Latam SaaS

Short answer (60 seconds): Anthropic launched Claude Opus 5 on July 24, 2026 at USD 5 / 25 per million tokens (input/output) — the same price as Opus 4.8, half the price of Fable 5. It leads agentic coding benchmarks with CursorBench and GDPval-AA at the top (Anthropic source) and ships with Fast mode 2.5x faster at 2x the base price ($10/$50) (Anthropic source). Versus Sonnet 5 ($2/$10 intro) it is 2.5x more expensive. The question for Latam SaaS founders is not "is Opus 5 better than Sonnet 5?" — of course it is. The question is when is it worth paying 2.5x more? Here is the Sonnet 5 / Opus 4.8 / Opus 5 / Fable 5 decision table, with real math and the operational checklist for this week.

A few days ago I wrote about Sonnet 5 dropping to $2/$10 and the takeaway was "audit your inference COGS this week". Today the question flips: for some workloads, Sonnet 5 is no longer enough and migrating to Opus 5 is worth the 2.5x. This post is the operational guide to decide which.

Disclosure: I have been testing Opus 5 since the API opened on July 24. The pricing and benchmark numbers come from official Anthropic, TechCrunch, Axios and The Verge publications, crossed with my own benchmarks on prompts representative of Latam SaaS products. Where I cite a figure without exact source (typically throughput, latency, my own measurements) I flag it inline.

What happened this week (and why it matters)

Anthropic launched Claude Opus 5 on July 24, 2026 as the new flagship of the Opus tier (Anthropic source, TechCrunch). Four launch facts that matter for a SaaS CFO/CTO:

1. Same price as Opus 4.8: $5 / $25 per million tokens. Anthropic did not raise the price to ship the improvements — you get better agentic coding benchmarks (CursorBench, GDPval-AA) and default thinking for free. It is half the price of Fable 5 ($10/$50 MTok), the family's top model. For a SaaS that was already on Opus 4.8, the upgrade is a no-brainer on capability; the only cost is the model ID migration in code.

2. Fast mode 2.5x faster, at 2x the base price. The least-covered note from the launch (Anthropic, The Verge): Opus 5 runs in accelerated mode 2.5x faster than standard mode, but it costs double: USD 10 / 50 per million tokens instead of $5 / $25. You pay 2x per token and finish the task in 0.4x the wall-clock — for interactive workloads (chat, agents in a loop with a user waiting) the cost per wall-clock minute can be equal or lower than Sonnet 5. It is the pricing trick that is not in the headline.

3. No 30-day data retention, unlike Fable 5. Opus 5 inherits Anthropic's zero data retention policy for API inference. Fable 5 retains data 30 days by default (needed for its personalized training features). For SaaS processing PII, financial or health data, Opus 5 is strictly better than Fable 5 on compliance. If your enterprise client asks you to "not store my prompts" and you are on Fable 5 today, migrating to Opus 5 unblocks that RFP.

4. More cybersec safeguards than Opus 4.8. Less visible but important: per The Verge, Anthropic gave Opus 5 more cybersec safeguards than Opus 4.8 — Opus 5 sits between Sonnet 5 and Fable 5 on cybersecurity capability, without reaching Fable 5. For security, legal or dual-use products that means Opus 5 will reject more borderline prompts than Opus 4.8 but fewer than Fable 5 — and never at Fable 5's level on pure cybersecurity tasks. If you sell into this segment, Opus 5 may not be your final model.

The macro context: last week we covered Anthropic's USD 5B AMD deal to diversify hardware. Opus 5 is the first product delivery on that strategy — the stable $5/$25 price is the promise that cheaper compute will reach production, not yet translated into lower customer pricing.

Decision table: Sonnet 5 / Opus 4.8 / Opus 5 / Fable 5

This is the table I wished I had the week of July 24. Pricing effective August 2, 2026 in USD per million tokens:

Row-by-row reading:

• Haiku 4.5 vs Sonnet 5: 2-2.5x cheaper. If your task is binary classification, JSON extraction, or 1-2 paragraph summarization, Haiku is enough. 80% of a typical Latam SaaS's traffic is this kind of task — the immediate savings are there, without thinking about Opus 5.
• Sonnet 5 vs Opus 5: Sonnet 5 costs 2.5x less. The capability gap is real (several points on CursorBench / GDPval-AA) but only matters in workloads where that gap changes business outcomes — agentic coding, long refactors, RAG over codebases, agents navigating repos.
• Opus 5 vs Opus 4.8: same price, improvement on CursorBench / GDPval-AA / ARC-AGI-3 per Anthropic. If you are on Opus 4.8 the upgrade to Opus 5 is free on capability — only the model ID changes. Anthropic did not raise the price to ship the improvements, which is atypical in the space.
• Opus 5 vs Fable 5: Opus 5 costs half. Fable 5 still wins on pure reasoning in very long sessions (4+ hours) and on workloads needing 30-day data retention. For 95% of SaaS, Opus 5 is the sweet spot.
• Opus 5 vs M3 / Kimi K3 (open weight): Opus 5 wins on managed API, ecosystem and governance. Open weight wins on self-hosting, local compliance and cost at high volume (>50M tokens/month). The decision is vendor lock-in vs control.

The number almost nobody looks at: Fast mode. Opus 5 in Fast mode runs 2.5x faster at 2x the base price ($10/$50 instead of $5/$25; Anthropic source). In my tests that drops p50 latency from ~3.2s to ~1.3s on 500-token generation. For interactive products (chat, agent in a loop with a user waiting) that changes UX. The "2x more expensive per token than standard Opus 5" becomes "cost similar to Sonnet 5 per wall-clock minute" in workloads where latency monetizes.

What "Opus 5 at $5/$25" means for your unit economics

The headline number sounds expensive until you do the math with your real workload. Here is the calculation across two typical Latam SaaS cases.

Base case: B2B SaaS in Spanish, 10,000 operations/month from an agent that generates rich content. Each operation: 2,200 input + 800 output (Sonnet 5's new tokenizer, +35% vs previous — Opus 5 inherits the same tokenizer).

Cost per 1 million combined tokens

Without cache, at the 73/27 input/output ratio typical for an agent:

Sonnet 5 at 1M tokens is 2.5x cheaper than Opus 5. Fable 5 is 2x more expensive than Opus 5 at the same ratio. The 73/27 is the average for an agent; if your workload is more output-heavy (document generation) the gap closes.

Monthly cost per 1,000 agentic operations (no cache)

Applied to the base case (3,000 tokens/op, 1,000 op = 3M tokens):

Opus 5 vs Sonnet 5 difference at 1M operations/month: USD 18,720/month, or USD 224,640/year. That is real money — but the question is whether those extra points on agentic coding benchmarks change outcomes for your case.

The case where Opus 5 is worth more than Sonnet 5 (whatever the cost)

Here is the trick Fast mode enables. If your agent does 5 internal iterations (tool calls + reasoning) before returning the final answer, Sonnet 5 takes ~12 seconds total; Opus 5 with Fast mode takes ~5 seconds — but at 2x the base price. In my tests Fast mode drops p50 from ~3.2s to ~1.3s on 500-token generation. If your product is interactive and lower latency improves conversion or retention, Opus 5 with Fast mode can be the right call even though it costs more per token: you pay 2x the base price ($10/$50 vs $5/$25) and finish the task in 0.4x the wall-clock.

Operational heuristic: if your workload is batch (runs overnight, no user watching), Fast mode does not help — you pay 2x the base price for speed you do not monetize. If it is interactive, the shorter wall-clock can more than pay the difference — but the unit economics calculation changes: it is no longer "Opus 5 vs Sonnet 5 at the same wall-clock" but "Opus 5 Fast vs standard Sonnet 5 at a different wall-clock".

Calculation disclaimer: the ratios assume a ~600-token system prompt, Spanish with ~38% tokenization overhead over English. If your prompt is shorter or the workload is different, run your own calculation.

What to audit in your SaaS this week

Operational checklist to run Monday to Friday. The goal is to end the week with evidence for planning, not intuition.

1. What percentage of your calls are multi-step agentic? If your traffic is <20% agentic (the rest is chat completion, classification, one-shot RAG), Sonnet 5 covers 80% and Opus 5 is for the other 20%. If your traffic is >60% agentic, Opus 5 with Fast mode can be the default choice. The metrics you need: average session length (in tool turns) and rate of retries / escalations.

2. Do you have data retention constraints? If your enterprise client requires 30-day data retention by contract, you are forced to Fable 5. If they require "explicit zero data retention" and you are on Fable 5 today, migrate to Opus 5. If you have no explicit constraint, Opus 5's default is already zero-retention — but confirm it in the contract.

3. Have you tested Fast mode? If your product is interactive and you have not tested Opus 5 with Fast mode, you are missing the most important data of the week. The calculation in the previous section assumes 2.5x faster; your real workload can be more or less. 30 minutes with the Anthropic playground and a representative prompt clears that up.

4. What is your cost per product feature, not per API call? The pattern I see fail most is looking at the aggregate Anthropic invoice without breaking it down by feature. What you need: a table with feature, inference cost, usage frequency, and the user plan that triggers it. Without that, the "Opus 5 yes / no" decision is made blind.

5. Do you have an A/B pilot running or planned? Do not migrate 100% of traffic blindly. Take 10-20% of agentic calls, route them to Opus 5, and measure (a) cost, (b) quality with your eval suite, (c) p50/p95 latency with Fast mode. One week of piloting gives you the answer that six months of debate do not.

When NOT to migrate to Opus 5

Opus 5 is competitive but not a universal replacement. Three cases where you should stay where you are:

1. Simple chat completions and one-shot summarization. If your product does "summarize this email", "translate this paragraph", "classify this ticket", Sonnet 5 performs well and costs 2.5x less. Opus 5 gives you capacity you do not use. The most common mistake in July 2026 was migrating simple workloads to Opus 5 "because it is newer". Do not fall for it.

2. Workloads with compliance that requires 30-day data retention. If your enterprise client requires 30-day data retention by contract (typical in financial services, health, government), Fable 5 is the only top-tier option. Opus 5 explicitly does not offer it. There is no upgrade path; you must stay on Fable 5 or migrate to a model that does support it.

3. High volume in Spanish with tight margins. If your SaaS runs at >10M tokens/month in Spanish and your gross margin is <40%, the 2.5x gap between Sonnet 5 and Opus 5 can be the line between profitable and break-even. The upgrade to Opus 5 is worth it if your pricing captures the value (Pro/Premium plans, B2B with high ACV). If it does not, the upgrade pushes you negative.

When NOT to stay on Sonnet 5 "just in case": if your eval suite fails on 5-15% of agentic cases with Sonnet 5, and the failure generates human escalation or a support ticket, the cost of the failure probably exceeds the Sonnet 5 savings. Calculate cost of failure (human time, churn) vs inference delta. Many times Opus 5 comes out cheaper in real terms even though it costs 2.5x per token.

Three operational practices that apply wherever you are

Regardless of which model you choose, three optimizations cut the invoice 30-50% combined and apply whether you are on Sonnet 5, Opus 5 or Fable 5:

1. Turn on prompt caching before changing model. If you have not enabled it yet, you have a quick win available without touching the model. On Opus 5, cache read costs $0.50/MTok — 5-10x cheaper than fresh input. Most SaaS with system prompts >500 tokens sees 50-70% hit rates without further changes. It is the change with the best ROI-per-time available today.

2. Route by task type, not by single default. A support agent does not need Opus 5 to classify (Haiku 4.5 is enough) or to generate simple replies (Sonnet 5 is enough). Routing by task type drops the invoice 50-60% without affecting UX. The hard part is instrumenting which task triggered each call — without that, you cannot route.

3. Monitor cost per tenant, not just globally. The "all API cost on a single line" pattern hides which tenants are profitable and which generate negative margin. Once you have cost per tenant you can (a) cap outliers before they burn your invoice, (b) upgrade power users with justification, (c) charge a premium for expensive features only some use.

Conclusion

Opus 5 does not replace Sonnet 5 — it complements it. The right question is not "do I migrate everything to Opus 5?" but "which 20% of my traffic is worth 2.5x more, and which 80% stays on Sonnet 5?". The answer depends on your workload mix, your pricing and your eval suite, not the press release.

Three operational recommendations for this week:

1. Audit your agentic vs one-shot mix. If your traffic is <20% agentic, Sonnet 5 is your default. If it is >60% agentic and/or interactive, test Opus 5 with Fast mode in an A/B pilot. The metric you need: % of calls that are multi-step vs single-shot.
2. Resolve your data retention constraint before choosing tier. If you need 30-day retention, Fable 5. If you need explicit zero retention, Opus 5 (not Fable 5). If you have no constraint, Opus 5 default is already zero — but get the contractual SLA in writing.
3. Do not migrate blindly. Run an A/B pilot with 10-20% of your agentic traffic. Measure cost, quality (with your eval suite) and p50/p95 latency with Fast mode. One week of real data gives you the answer that six months of debate do not.

If your startup is running on Claude or figuring out how to split traffic between Sonnet 5, Opus 5 and Fable 5, and you want to validate the real impact on your numbers before committing, book a free 30-minute call — in 20 minutes we can usually calculate your monthly savings with your real workload mix.

---

Read also:

• Claude Sonnet 5 at $2/$10: inference cost dropped up to 80% — the companion post: when Sonnet 5 makes sense and how the savings were calculated.
• Model Context Protocol (MCP) explained for SaaS founders — the open standard that defines how an AI assistant connects to external tools, key if you are going to route between Sonnet 5, Opus 5 and Fable 5 in production.
• Claude Opus 5: what changes versus Opus 4.8 and when to use it — the complete technical guide to the model (specs, SDK, benchmarks, open-weight comparison).
• Anthropic signed USD 5B AMD deal — why Opus 5's price will hold steady (and eventually drop).
• Back to the blog — all articles.

Frequently asked questions

How much does Claude Opus 5 cost?

USD 5 per million input tokens and USD 25 per million output tokens, identical to Opus 4.8 pricing. It is half the price of Fable 5 ($10/$50 MTok) and 2.5x more expensive than Sonnet 5 during the introductory pricing ($2/$10 MTok, through August 31, 2026). Opus 5 does not have Sonnet 5's lower intro price — the $5/$25 is stable from launch on July 24, 2026.

Does Opus 5 replace Sonnet 5 or Opus 4.8?

Neither. Opus 5 occupies a different slot: it is "almost Fable 5 at half the price", not "more expensive Sonnet 5". Versus Sonnet 5 ($2/$10) it is 2.5x more expensive, but leads agentic coding benchmarks (CursorBench, GDPval-AA, ARC-AGI-3) per Anthropic's own page. Versus Opus 4.8 ($5/$25) it costs the same but improves the same benchmarks by roughly 5-10 points depending on the suite. The correct decision depends on workload: short tasks / high volume stay on Sonnet 5; long, complex, agentic tasks go to Opus 5; and only if you need the absolute ceiling do you pay for Fable 5.

What is Fast mode and how much faster is it?

Fast mode is Opus 5's accelerated inference option that runs the model 2.5x faster than standard mode, at 2x the base price: USD 10 input / USD 50 output per million tokens (double the standard $5/$25). Anthropic announced it on July 24, 2026. You pay 2x per token but you finish the task in 0.4x the wall-clock — for interactive workloads (chat, agents in a loop with a user waiting) the cost per wall-clock minute can be equal or lower than Sonnet 5. For offline batch, you pay 2x for speed you do not monetize.

Does Opus 5 retain my data?

No by default. Opus 5 inherits Anthropic's policy of not using API data to train models, and data retention is zero (not 30 days like Fable 5). For workloads with sensitive data (PII, financial, health) that is an improvement over Fable 5. If you need the explicit zero data retention policy under an enterprise contract, it is available via Anthropic's Zero Data Retention add-on — but the default is already strict.

When does it NOT make sense to migrate to Opus 5?

Three clear cases. (1) Simple chat completions or one-shot summarization where Sonnet 5 performs well and costs 2.5x less. (2) Classification or data extraction tasks where Haiku 4.5 ($1/$5) or Sonnet 5 ($2/$10) suffice. (3) Workloads where you need 30-day data retention for compliance (Fable 5 is still the only top-tier option that offers it — Opus 5 does NOT). In any of these three cases you pay 2.5x more for capacity you do not use.