Claude Sonnet 5 vs Opus 4.8 vs GPT-5.5: agent cost math

Anthropic launched Claude Sonnet 5 yesterday, 30 June 2026, at an introductory price of 2 dollars per million input tokens and 10 dollars per million output tokens, valid through 31 August. The company calls it the most agentic Sonnet model yet, and its own cost-performance curves show it matching Claude Opus 4.8 on some tasks at a fraction of the cost. The model is available from day one on the Claude API, AWS and Microsoft Foundry, with Google Vertex to follow. The same news cycle closed the month's biggest availability story: Anthropic has restored global access to Fable 5 and Mythos 5, ending the eighteen-day pause triggered by a US export-control directive on 12 June. For an Azure-first team in Sweden, both halves of the announcement matter. The first resets the default model choice for agent workloads. The second is a lesson in continuity planning you should not let go to waste.

What Anthropic shipped on 30 June

Claude Sonnet 5: the numbers that are public

Sonnet 5 posts 63.2 percent on SWE-bench Pro and 80.4 percent on Terminal-Bench 2.1, with broad improvements over Sonnet 4.6 in reasoning, tool use, coding and knowledge work. On the agentic evaluations Anthropic highlights, BrowseComp and OSWorld-Verified, the pitch is not a new absolute record but substantially improved cost efficiency at medium effort. That framing is deliberate. Anthropic is positioning Sonnet 5 as the model you run at volume, with Opus 4.8 reserved for the work that justifies its rate card.

Two details from the launch deserve attention from platform teams. First, the introductory pricing of 2/10 dollars per million tokens runs through 31 August 2026; the regular price is 3 dollars in and 15 dollars out. Budget on the regular price and treat the summer rate as a migration subsidy. Second, Sonnet 5 ships with cybersecurity safeguards enabled by default and is, per Anthropic's own system card language, substantially weaker at exploit development than the Opus line. Given what happened in June (more below), that is not an accident.

Fable 5 and Mythos 5 are back

On 12 June, a US export-control directive forced Anthropic to suspend access to Fable 5 and Mythos 5 after researchers at Amazon demonstrated a method to bypass safety controls and extract working exploit code for software vulnerabilities. Notably, the reporting indicates the behaviour was not unique to Anthropic: older models from multiple providers exhibited the same weakness. Anthropic's fix is an updated automated classifier that blocks the reported exploitation technique in more than 99 percent of trials, deliberately tuned with a wide safety margin that also catches ambiguous developer prompts. Federal security reviews completed, and access is restored as of this week.

If your workloads sat on Fable or Mythos, you just lived through an eighteen-day partial outage of a production dependency, imposed by a regulator in a third country, with no SLA remedy that covers it. File that experience; we return to it below.

Claude Science, briefly

Anthropic also used 30 June to launch Claude Science, a research product it positions the way Claude Code is positioned for software engineering: an environment where the model autonomously executes meaningful scientific work from high-level instructions, with tooling for genetics, chemistry and protein biology. It is available to all paid Claude subscribers. For most enterprise platform teams it is not this quarter's concern, but life-science and pharma organisations in the Nordics should put an evaluation on the calendar.

The price sheet as of 1 July 2026

Here is the comparison that actually matters for model routing decisions this quarter. GPT-5.6 was previewed to a small partner group last week but is not generally available, so the OpenAI column is GPT-5.5, the current production model.

Claude Sonnet 5: 2 dollars in / 10 dollars out per million tokens through 31 August, then 3 / 15.
Claude Opus 4.8 (released 28 May): 5 / 25, with a 2.5x-speed fast mode at 10 / 50.
GPT-5.5: 5 / 30.

Read the ratios, not the absolute numbers. At the regular rate, Sonnet 5 costs 60 percent of Opus 4.8 on both input and output. At the introductory rate it costs 40 percent. Against GPT-5.5 the gap is wider still on output tokens, which is where agent workloads bleed money: 10 dollars versus 30 at the summer rate. Anthropic's claim that Sonnet 5 approaches Opus 4.8 on many agentic tasks is a vendor claim and you should run your own evals, but the rate card alone shifts the burden of proof. The question is no longer "can we afford the frontier model everywhere?" It is "which specific tasks still justify it?"

Cost per task: a worked example

These figures are an illustration, not a benchmark. The token profile below is a modelling assumption you should replace with numbers from your own telemetry. The arithmetic is the point: it shows how the rate cards compound across a realistic multi-turn agent loop.

Assume a mid-complexity agent task: a support triage or document-processing agent that makes 12 model calls per task, averaging 5,000 input tokens per call (system prompt, accumulated context, tool results) and 500 output tokens per call. That is 60,000 input and 6,000 output tokens per completed task.

Per task (60k in / 6k out):

Sonnet 5 (intro 2/10):    60k x 2/1M  + 6k x 10/1M  = 0.12 + 0.06 = 0.18 USD
Sonnet 5 (regular 3/15):  60k x 3/1M  + 6k x 15/1M  = 0.18 + 0.09 = 0.27 USD
Opus 4.8 (5/25):          60k x 5/1M  + 6k x 25/1M  = 0.30 + 0.15 = 0.45 USD
Opus 4.8 fast (10/50):    60k x 10/1M + 6k x 50/1M  = 0.60 + 0.30 = 0.90 USD
GPT-5.5 (5/30):           60k x 5/1M  + 6k x 30/1M  = 0.30 + 0.18 = 0.48 USD

At 50,000 tasks per month:

Sonnet 5 (intro):    9,000 USD
Sonnet 5 (regular): 13,500 USD
Opus 4.8:           22,500 USD
GPT-5.5:            24,000 USD

At list prices, routing this workload to Sonnet 5 instead of Opus 4.8 saves 9,000 dollars a month, or 108,000 dollars a year, on a single workload of moderate volume. Two further levers stack on top for Anthropic models: prompt caching cuts the cost of cached input tokens by 90 percent, which bites hard in agent loops where the system prompt and tool definitions repeat on every call, and batch processing halves the price of anything that can tolerate asynchronous completion. An agent with a stable 3,000-token system preamble across 12 calls per task is exactly the shape caching was built for.

There is one honest complication in cost-per-task math: a weaker model that fails more often does not save money. If Sonnet 5 completes a task at 92 percent success and Opus 4.8 at 96 percent, and a failure costs you a retry, the effective cost gap narrows but nowhere near closes: a full retry on Sonnet still lands under half an Opus first attempt in the example above. Where failure cost is not a retry but a human escalation or a bad customer outcome, the calculation changes entirely. That is a routing question, not a pricing question.

When the cheap model is the right model

A useful default for July 2026: Sonnet 5 is the workhorse, Opus 4.8 is the escalation tier, and the router between them is a piece of engineering you own. The cheap model is the right model when the following conditions hold.

The output is verifiable. Code that must compile and pass tests, extractions validated against a schema, answers checked against a source document. When a cheap automated check catches failure, retrying a cheap model beats defaulting to an expensive one.
The tool surface is bounded. Agents with a handful of well-described tools and short decision horizons are where Sonnet-class models have closed most of the gap. Long-horizon, open-ended work with dozens of tools still rewards the frontier tier.
A retry is cheaper than the premium. Do the arithmetic above with your own failure rates. If cost of failure is a retry, Sonnet wins at almost any plausible success-rate gap. If cost of failure is a compliance incident, it does not.
You have evals in place. Without a task-level evaluation suite, you cannot measure the success-rate gap, so you cannot defend either choice. Build the eval before the router.
Latency budgets are ordinary. Opus 4.8's fast mode exists for the cases where a frontier answer must also be quick, at double the frontier price. If you are reaching for it by default, your routing is wrong.

Invert each condition to find the Opus 4.8 workloads: unverifiable high-stakes output, sprawling tool surfaces, long autonomous sessions, and tasks where one bad completion costs more than a month of model savings. In our experience those are a minority of enterprise agent traffic by volume, which is precisely why the Sonnet 5 rate card matters more than the Opus benchmark table for most budgets.

A practical migration note for Azure-first teams: Sonnet 5 is on Microsoft Foundry at launch. That means the evaluation can run inside your existing Azure subscription, under your existing enterprise agreement and network controls, without a new procurement exercise. Pin the exact model version in your deployment configuration, run your eval suite against your top three agent workloads, and compare completed-task cost, not per-token cost, against whatever you run today.

The eighteen-day pause is the other half of the story

The Fable and Mythos suspension deserves more attention than it will get now that access is restored. Consider what actually happened: a security research finding, an export-control directive from the US government, and within days two production model families became unavailable worldwide, including to European customers with no US nexus at all. It lasted eighteen days. No amount of Azure region redundancy would have helped, because the dependency that failed was legal, not technical.

The engineering response is unglamorous and well understood: put a model-abstraction layer between your agents and any single provider, keep a tested fallback configuration for each critical workload, and rehearse the switch the way you rehearse a region failover. The June pause gives you the incident report to justify the work. A fallback that maps each Opus workload to Sonnet 5, or to a model from a second provider, with known eval scores on your own tasks, converts a future eighteen-day incident into a config change and a quality note.

The Swedish and EU angle

Residency and hosting. Day-one availability on Microsoft Foundry is the headline for Swedish enterprises, but availability on a platform is not the same as availability in your region. Before you promise data residency to a DPO, verify which Foundry regions and data-processing options actually serve Sonnet 5, and document the flow the same way you did for your current models. Do the verification now, while the routing decision is still open.

AI Act housekeeping. Transparency obligations for general-purpose AI models have applied since August 2025, and the next major application milestone arrives on 2 August 2026. Swapping a model inside an agent system is exactly the kind of change your internal AI governance should capture: update the model inventory, re-run the evaluations you cite in your documentation, and record the routing logic. A model swap that saves 40 percent and is documented in an afternoon is a good trade. An undocumented one is technical debt with a regulatory flavour.

Procurement and continuity. The export-control pause is a concrete, citable example of third-country regulatory risk to bring into supplier assessments and continuity planning, the kind of scenario NIS2-driven risk work keeps asking for. Ask your model providers, through whichever cloud you buy them, what their notification and remediation commitments are when access is suspended for regulatory rather than technical reasons. For budgeting, use the post-August list price of 3/15 dollars, treat the introductory rate as upside, and note the asymmetry the June incident revealed: pricing is contractual, but availability can be overridden from outside the contract.

Decision checklist

1. Pull your telemetry. Get real input/output token profiles per completed task for your top three agent workloads. Redo the cost math above with your numbers at 2/10, 3/15, 5/25 and 5/30.
2. Run Sonnet 5 through your eval suite this month. The introductory pricing window through 31 August is a cheap evaluation subsidy. Measure completed-task success rate against your current model, not benchmark deltas.
3. Build or fix the router. Default to Sonnet 5 where output is verifiable and retries are cheap; escalate to Opus 4.8 on the workloads that fail the checklist above. Log every routing decision.
4. Stack the discounts. Restructure prompts so static preambles cache, and move non-interactive workloads to batch. These compound with the cheaper model.
5. Verify Foundry region coverage for Sonnet 5 before committing residency-sensitive workloads, and update your Article 30-adjacent data-flow documentation when you switch.
6. Write the fallback plan. Map every production model to a tested alternative, note the eval delta, and rehearse the switch. The eighteen-day Fable pause is your justification slide.
7. Budget at list price. Plan on 3/15 from September; treat anything better as margin.