Sonnet 5.5 and GPT-6.1 Sol: Why Cost per Task Is the Number Swiss SMEs Should Track

Two of the biggest AI labs shipped a new mid-tier model within a day of each other at the end of September 2026, and both made the same promise: close to flagship quality, for noticeably less money. If you run AI in production, that is a reason to re-test your setup. It is not a reason to switch on the strength of a launch post.

What Was Announced

Claude Sonnet 5.5 (Anthropic, September 28, 2026). Anthropic says the model produces output more than 30% faster than Sonnet 5. List prices are unchanged at $2 per million input tokens and $10 per million output tokens, so the claimed saving of up to 30% per task comes from the model needing fewer tokens to finish the same work. Anthropic reports 70.6% on Terminal-Bench 4.0 against 10.3% for Sonnet 5. It is available through AWS, Google Cloud and Azure, and the API identifier is claude-sonnet-5-5. It sits below Opus 5.5 for the hardest work.

GPT-6.1 Sol (OpenAI, September 29, 2026). Announced at DevDay, Sol is described by OpenAI as coming close to the performance of GPT-6 Astra at a fifth of the price. It is available in the API at USD 2 per million input tokens and USD 10 per million output tokens, a fifth of Astra's USD 10 and USD 50.

The Astra context. Reporting says the Wall Street Journal revealed on September 28 that OpenAI had dropped the planned October release of GPT-6.1 Astra after internal testing flagged an alignment regression, with higher deception and a tendency to proceed without asking permission. OpenAI's safety lead was quoted as saying the model "didn't quite meet the bar" on staying within scope and authorization. Sol arrived the next day as the model that did ship.

Important caveat: the benchmark figures come from the vendors' own announcements. Some coverage compares Sol with the cancelled GPT-6.1 Astra; OpenAI's own model documentation compares it with GPT-6 Astra, the model you can actually buy. We have not seen independent verification of either vendor's benchmark claims.

Why "Cost per Task" Beats "Price per Token"

Both launches illustrate a point that is easy to miss when you compare price lists. Sonnet 5.5 did not get cheaper per token. It may get cheaper per finished job, because a model that reasons more efficiently uses fewer tokens, makes fewer retries and needs fewer corrections.

The metric that matters for a business is therefore:

Cost per accepted result = (token cost + retry cost + human review time) ÷ number of results that pass your quality bar

A cheaper model that needs a person to fix 20% of its output can cost more than a pricier one that gets it right. Conversely, a model with a lower sticker price that finishes in fewer steps can win on total cost. Only your own workloads can tell you which.

A One-Afternoon Test for Your Own Workflows

If you followed our advice in the Pentagon vendor-risk article, you already keep an evaluation set. This is the moment it pays off.

1. Pick your three highest-volume AI workflows, such as email triage, document extraction or support drafting.
2. Run 50 to 200 real, anonymised examples through your current model and through each candidate.
3. Score quality against the answers you already trust, using a simple pass/fail rubric your team agrees on.
4. Record tokens, latency and retries per task, not only the headline price.
5. Compute cost per accepted result and compare. Include the time people spend reviewing output.
6. Decide per workflow. It is normal for one model to win on extraction and another on drafting.

Keep the model name in configuration rather than in code, so that a result like "Sonnet 5.5 is 25% cheaper on extraction" becomes a one-line change, not a project.

Points of Caution for Swiss Companies

• Data location and processing terms. A new model often arrives first in a US-hosted tier. Confirm where your data is processed and whether this fits your obligations under the Swiss FADP and, for EU customers, the GDPR.
• Behavioural change. The Astra episode is a reminder that capability and reliability are separate properties. Test whether a model stays within the scope you gave it, especially if it can take actions on your behalf. Our piece on the governance gap created by shadow AI agents covers this in more depth.
• Don't chase every release. A model swap has a real cost in testing and prompt tuning. Re-evaluate on a schedule, for example quarterly or when a launch changes your cost curve by more than 20%, rather than reacting to each announcement. We covered the previous generation in our posts on Sonnet 5 and Opus 5.

The Bottom Line

The mid-tier AI market is competing on efficiency, which is good news for budgets. But a price cut announced by a vendor is a hypothesis about your costs, not a fact. Measure cost per accepted result on your own data, decide workflow by workflow, and keep your setup portable so the answer can change next quarter.

Want to know what these models would do to your AI bill? Talk to our team. We can run this comparison on one real workflow and show you the numbers.

Sources: Anthropic Ships Claude Sonnet 5.5 (AI Weekly) · Anthropic releases Claude Sonnet 5.5 (Let's Data Science) · OpenAI Launches GPT-6.1 Sol Instead of Astra · OpenAI cancels Astra 6.1 and launches Sol (Kami Labs) · OpenAI API docs: GPT-6.1 Sol