GPT-6 Sol vs. Claude Opus 5.5 for Coding: What Changed for Internal Builders

For raw coding capability, Claude Opus 5.5 wins outright. For cost per task up to a real ceiling, GPT-6 Sol wins, and wins by a lot. That's the short answer to which model to reach for in late 2026. The longer answer matters more, because it's already changing how internal tools get built.

On September 22, 2026, Anthropic shipped Claude Opus 5.5. About 90 minutes later, OpenAI shipped GPT-6 Sol and its smaller sibling GPT-6 Luna.¹² Coincidence or not, the timing forced a direct comparison that matters for anyone deciding whether to renew a SaaS seat or have someone on the team build the tool instead.

The two-model launch everyone's asking about

Anthropic's pitch for Opus 5.5 is capability at a lower price: the company says it matches its larger Claude Fable 5.1 on most work while costing 40% less to run than the previous Opus 5.³ OpenAI's pitch for Sol and Luna leads with cost instead: both models are priced 50% below their GPT-5.6 predecessors' promotional rates, which OpenAI credits to caching and inference efficiency gains.²⁴

Those framings aren't marketing spin. They describe where each lab is actually optimizing. Anthropic is defending the top of the capability curve. OpenAI is racing down the cost curve. Both moves land on the same desk: the person deciding whether to renew a SaaS contract or build the tool in-house this quarter.

The benchmark verdict, in plain English

Strip away the marketing and look at Artificial Analysis's independent Intelligence Index, which blends coding and agentic benchmarks into one score. At maximum reasoning effort, Claude Opus 5.5 scores 58. GPT-6 Sol scores 48.⁵ Opus 5.5 leads, and it isn't close.

But the index score hides the more useful number: cost per task. Up to an index score of roughly 44, GPT-6 Sol is the cheaper way to get there.⁶ Push past that point and Sol runs out of headroom. Its best possible result, 47.5 at max effort, costs $1.06 per task and still scores below what Opus 5.5 delivers at its default medium-effort setting: 51.2 points for $1.34 per task.⁵⁶ That's 3.7 more intelligence points for 26% more money. On Terminal-Bench 4.0, a benchmark built around real agentic coding tasks, Opus 5.5 at medium effort scores 52.5% against Sol's 43.9% at max effort.⁶

The practical takeaway: Sol is the value play for tasks that don't need the top tier of reasoning. Opus 5.5 is the model you reach for when the task is genuinely hard and an extra dollar per run is trivial next to the cost of getting it wrong. Anthropic's own testing backs that up. Opus 5.5 fixed load-time inefficiencies across a web app correctly 39 out of 40 times, while the older Opus 5 made smaller fixes that also altered the app's behavior.³ One early tester used it to migrate 680,000 lines of code in under a day, work that would have taken an engineering team weeks.³

Is GPT-6 Sol actually half the price of Claude Opus 5.5?

Here's the sticker price breakdown, because the cost gap isn't as simple as "half off":

• Claude Opus 5.5: $4 per million input tokens, $20 per million output tokens, a 20% cut from Opus 5's $5/$25. Cache reads dropped 60%, to $0.20 per million tokens.³⁷
• GPT-6 Sol: $2 per million input tokens, $10 per million output tokens, down 50% from GPT-5.6 Sol's promotional pricing.²⁴
• GPT-6 Luna: $0.10 per million input tokens, $0.50 per million output tokens, also down 50%.²⁴

On paper, Sol looks like half the price of Opus 5.5. In practice it isn't, because output tokens dominate real coding costs, and Sol tends to need more of them to reach a comparable result. Opus 5.5 at max effort produces around 119,000 output tokens per task versus Sol's 31,000 at max effort.⁵ A model that writes less code to solve the same problem can end up cheaper per task even at a higher per-token rate. That's the trap in reading model pricing as a simple ratio: cache behavior, context length, and how verbose a model is under the hood all move the real bill more than the sticker price does.

Anthropic claims typical Opus 5.5 workloads cost 40% less than the same workloads on Opus 5, thanks to the pricing cut plus output that generates over 30% faster.³ OpenAI, for its part, says GPT-6 Sol makes about half as many factual mistakes as its predecessor on internal evaluations, which matters for anyone counting the cost of a bad code suggestion, not just the token bill.¹

Why this matters for build vs. buy, not just coding benchmarks

Benchmarks are an engineering story. Build vs. buy is an accounting story, and the numbers there are getting hard to ignore.

The average enterprise now runs about 305 SaaS applications and spends $55.7 million a year on SaaS, up 8% year over year even as the number of apps has essentially flattened.⁸ That growth is coming from AI-driven pricing changes inside existing subscriptions, not new app sprawl: 78% of IT leaders reported unexpected charges tied to AI features or consumption-based pricing.⁸ As Internal AI Workflows Cost Pennies. SaaS Subscriptions Still Cost $9,324 a Seat. lays out, the gap between per-token inference cost and per-seat licensing is only widening.

Meanwhile the cost of writing internal software with a frontier model just dropped again, on both sides of the lab race. A workflow automation, an admin dashboard, an internal approval tool: the kind of thing a team used to pay $200 a month in SaaS fees for now often gets built for a few dollars in tokens, as we've covered in how AI frontend generation is killing niche SaaS tools. When the token cost of a build is a rounding error next to a year of seat licenses, the build-vs-buy decision stops being about ideology and starts being arithmetic, which is exactly the shift covered in the Fable 5.1 build-vs-buy math.

The shadow AI reality check

This isn't hypothetical. It's already happening, mostly without anyone's permission.

In a Retool survey of 817 builders and customers, 35% of organizations have already replaced at least one SaaS tool with a custom-built alternative, and 78% expect to build more custom internal tools in 2026.⁹ Workflow automations and internal admin tools are the categories builders say are most at risk of replacement, ahead of BI tools, CRMs, and project management software.⁹

More striking: 60% of builders say they've built software, workflows, or automation outside official IT oversight in the past year, and a quarter do so frequently, citing speed, unmet needs, and slow IT process as the top reasons for going around IT.⁹ That lines up with broader shadow AI numbers. Independent surveys converge on roughly 55% to 78% of employees using AI tools at work that were never sanctioned by their employer.¹⁰ Cheaper, more capable models don't create this behavior. They accelerate it, and they raise the governance stakes for whoever eventually has to account for what got built.

When should you build instead of buy?

The two-model launch gives builders a cleaner cost ladder than they had a month ago. Use it like this:

1. High-volume, low-complexity automations. Route these to Sol-class or Luna-class pricing. If the task is a repeated internal workflow, an intake form, a status report generator, marginal cost matters more than squeezing out the last few points of intelligence.
2. Complex, one-off builds. Reserve Opus-class spend for the migration, the gnarly refactor, or the tool where a wrong answer is expensive. The extra dollar per task is nothing next to a broken production system.
3. Anything compliance-heavy or deeply integrated. Keep buying SaaS here. Model cost isn't the bottleneck when the real cost is audit trails, vendor certifications, or a dozen existing integrations that would take longer to rebuild than to keep paying for.
4. Anything with a semi-technical builder and a clear spec. This is the sweet spot the pricing shift actually opens up. A person who can write a clear requirement and check the output, paired with either model, now regularly beats a SaaS seat on pure cost, especially over a multi-year contract.

A useful gut check, borrowed from how The End of the Large Software Team frames it: if the SaaS tool in question is niche, single-purpose, and mostly used by one team, the odds that building beats buying just went up again. If it's a platform with network effects, deep integrations, or regulatory weight, the calculus barely moved.

If your team is testing this, a platform like Remy is built for exactly this middle ground: giving semi-technical builders a fast path from spec to running internal tool without needing a dedicated engineering team to own the result.

What to watch next

Anthropic has already signaled this isn't the last price cut. Opus 5.5 arrived just two months after Opus 5, and Anthropic has described it as its first release since CEO Dario Amodei embraced calls to pace the frontier rather than sprint it.⁷ Expect cheaper Sonnet and Haiku tiers to follow the same pattern Sol and Luna just set: same capability, lower price, faster.

The bigger question isn't which model wins the next benchmark. It's whether governance keeps pace with how fast building is becoming the default. With 51% of builders already shipping AI-built production software their teams actually use, and 35% of organizations still running with no AI productivity metrics at all,⁹ the gap between what gets built and what gets tracked is where the real risk sits going into 2026.