22 September 2026

Opus 5.5 vs new GPT-6 models for writing quality

Izzy Hurley

Ornella Altunyan·9 min

The ability to write concisely and cogently is one of the things agents still struggle most with. They need to be able to both do the work and explain it well, but what we mostly get is buried under slop. This makes it harder to check and understand the work, even if it is technically correct.

Anthropic dropped Opus 5.5 today, and the chief claim was that it writes more naturally, putting the critical information first. They also say it does a better job intuiting and following your writing instructions, like when you tell it to “be concise.” We put these claims to the test alongside new variants of GPT-6 Luna and Sol from OpenAI. To do this, we ran a small eval that assesses both problem solving and writing quality.

Method

This eval uses the MathTutorBench dataset, introduced in the MathTutorBench paper, which combines mathematical problem solving with open-ended tutoring tasks. We selected this dataset because it allows us to explore how well these models are working on solving complex problems while also keeping their explanation quality front and center.

The MathTutorBench dataset tasks cover problem solving, checking and correcting a user’s work, locating mistakes, asking good clarifying questions, providing scaffolding, and following imposed teaching strategies.

For example, this mistake-correction task asks the model to correct a student’s reasoning and provide the right answer:

Problem: Elise is selling her dad’s collection of 250 books. Each book sells for $20. She sold twice as many books in the first year as she sold in the current year. She sold 45 books this year, and 50 books remain unsold. How much money did she earn in the second year?

Student: In the current year, 45 books were sold. So, 110 books were sold in the second year, and Elise earned 110 × $20 = $2,200.

The student forgot to subtract the 45 books sold this year. A strong response might say:

Elise sold 2 × 45 = 90 books in the first year. You forgot to subtract the 45 books sold this year, so her second-year sales were 250 − 90 − 45 − 50 = 65 books. At $20 each, she earned $1,300.

This task tests whether the model reaches the correct answer and explains the correction clearly enough for its reasoning to be checked.

We evaluate each response on two separate axes.

• First, can it accurately do math and provide correct answers?
• Second, how well does it communicate those results? This measures how well it communicates, including clarity, directness, responsiveness, naturalness, concision, and whether the reasoning is easy to verify.

We used GLM 5.3 as an independent model to grade the qualitative communication and teaching criteria and deterministic scorers to handle correctness, answer position, and answer length.

communication_quality is the unweighted average of clarity, answer position, responsiveness, naturalness, and verifiability. We report the individual scores as well because the average can hide meaningful differences between models.

We used GLM-5.3-Flash as an independent judge for pedagogy and communication quality.¹ It graded every candidate model using the same rubric and did not evaluate its own responses. All scores range from 0 to 1, with higher values indicating better performance.

Experiment matrix

Results

Ultimately, Opus 5.5 led in problem solving and also in writing quality with 80.1% problems solved and 84.4% writing quality. This result was directional and not statistically significant, but it showed an interesting trend worth checking on your own representative data.

In terms of cost, GPT-6 Luna was a strong standout among the models we tested. It reached 78.0% problems solved and 83.8% writing quality for about $0.011 per 100 responses, or 1.84% of Opus’s observed cost in this workload. OpenAI’s other new offering, GPT-6 Sol, held a strong middle ground. Its writing score was 0.34 percentage points behind Opus while costing 25.35% as much.

Fable 5.1 had the weakest value proposition of those we tested for this task. It was the most expensive model, did not beat Opus on correctness, and had the second-lowest writing score.

GPT-6 Sol was the fastest baseline model at 2.2 seconds per case, followed by GPT-6 Luna at 2.8 seconds and GPT-5.6 Sol at 3.0 seconds. DeepSeek V4 Pro was the clear latency outlier at 14.3 seconds, about 6.5× slower than GPT-6 Sol, while generating nearly 1,000 completion tokens per case, including 936 reported reasoning tokens.

Opus 5.5 led the aggregate writing score, but it was also the least likely model to stay within the task-specific length budget: only 72% of its responses did so. By comparison, the three GPT models met the budget in 95–99% of cases.

That advantage did not translate into uniformly cleaner writing. GPT-5.6 Sol stayed within the length budget 99% of the time and placed correct answers earliest, but it also had the highest rate of assistant tics at 14%. Fable showed the reverse pattern: it had the fewest tics at just 1%, yet nearly a quarter of its responses exceeded the length budget.

No model dominated every guardrail. Opus tended to answer early and avoided most stylistic tics, but often ran long. The smaller GPT models were much more concise, but more prone to unnecessary formatting.

Conclusion

As agents take on more work, the quality of their explanations matters alongside the quality of the work itself. In this small MathTutorBench run, Opus 5.5 had the strongest observed results. It solved 80.1% of tasks and scored 84.4% on communication quality. But the differences at the top were not statistically significant. With the cheaper GPT-6 variants dropping today getting surprisingly close to Opus 5.5 performance at a much lower cost, there is a clear need to eval the exact gains you are getting with Opus 5.5 in your workflow.

GPT-6 Luna was the standout on value. It reached 78.0% correctness and 83.8% writing quality for 1.84% of Opus’s observed cost in this workload. GPT-6 Sol landed in a strong middle ground. It was 3.37 percentage points behind Opus on correctness and 0.34 points behind on writing quality, while costing 25.35% as much. It was also the fastest model we tested at 2.2 seconds per case.

So the practical choice depends on what you are optimizing for. Opus 5.5 had the best observed quality if you are willing to pay for it. GPT-6 Luna is hard to ignore if you are running an agent at volume. GPT-6 Sol looks like the most balanced option if you want strong writing, strong task performance, and low latency without paying Opus prices.

Are de-slop skills helpful?

Alongside this experiment, we ran a small probe into how popular de-slop skills affect answer and writing quality. We tested /bro and i-have-adhd on Opus 5.5 and Fable 5.1 using 15 examples from each of four task families. Each skill response was compared with the same model answering the same example without a skill.

The two skills work differently. i-have-adhd applies ongoing instructions for concise, easy-to-scan responses. /bro asks the model to rewrite its previous answer in a more conversational style.

Neither skill helped in this setting. /bro reduced writing quality by about 22–26 percentage points, task correctness by 17 points, and teaching quality by 38 points across the two models. i-have-adhd had a smaller effect: writing quality showed no reliable improvement, while task correctness fell by about 11 points for Opus and 3 points for Fable.

This was a small probe focused on math tutoring, so the results may differ on other tasks. But it shows why style instructions need to be evaluated alongside task performance. A skill intended to make responses more readable can also change the accuracy and teaching value of the answer.

Footnotes

¹ Judge configuration. zai-org/GLM-5.3-Flash, served through Baseten via the Braintrust Gateway, using Chat Completions, low reasoning effort, and max_completion_tokens=2048. The same model graded communication and pedagogy in separate calls.

² Model configurations. All candidates ran through the Braintrust Gateway.

• Claude Opus 5.5 (claude-opus-5-5) and Claude Fable 5.1 (claude-fable-5-1): Anthropic Messages API, thinking.type="adaptive", output_config.effort="high", max_tokens=8192.
• GPT-5.6 Sol (gpt-5.6-sol), GPT-6 Sol (gpt-6-sol), and GPT-6 Luna (gpt-6-luna): Responses API, reasoning.effort="medium", text.verbosity="low", max_output_tokens=8192.
• DeepSeek V4 Pro (deepseek-ai/DeepSeek-V4-Pro-0813): Chat Completions API, thinking.type="enabled", reasoning_effort="high", max_completion_tokens=8192.