New models tested: Fable 5.2, Muse Ember 1.0, Inkling 2 Small
Summary
I tested several new models: Fable 5.2, Muse Ember 1.0, and Inkling 2 Small, across the same set of five workloads.
Fable 5.2 is our new leader on design, but it is very expensive. Muse Ember surprised me: it is close to Sol 6 on overall quality at a much lower cost. I recommend trying it. Inkling 2 Small is competitive on quality, but slow. Great to see a new competitor on the field.
Fable 5.2 and Fable 5.1
Fable 5.2 is a big improvement over Fable 5.1 and our new leader on design. It also scored higher on coding, while extraction was nearly unchanged. But frontier intelligence is getting more expensive faster than it is getting better.
| Model | Mean quality | Cost / task | Latency |
|---|---|---|---|
| Fable 5.2 | 0.884 | $0.4279 | 77.2 s |
| Fable 5.1 | 0.806 | $0.2833 | 44.6 s |
Quality by workload
| Model | Extraction | Matching | Research | Design | Coding |
|---|---|---|---|---|---|
| Fable 5.2 | 0.935 | 0.981 | 0.986 | 0.731 | 0.787 |
| Fable 5.1 | 0.939 | 0.975 | 0.963 | 0.471 | 0.684 |
My takeaway: A new leader on design, but frontier intelligence is getting more and more expensive.
Fun fact
Asked: “This answer is garbage. Fix it.”
Fable: “I skip the defensiveness and the groveling.”
Muse Ember 1.0
Muse Ember does really well across a wide range of tasks. It tied for the highest research score, was near the top on extraction, and did well on coding, with quality close to Sol 6 at a much lower cost.
| Model | Mean quality | Cost / task | Latency |
|---|---|---|---|
| Muse Ember 1.0 | 0.880 | $0.0053 | 16.1 s |
| Sol 6 | 0.881 | $0.0291 | 12.2 s |
Quality by workload
| Model | Extraction | Matching | Research | Design | Coding |
|---|---|---|---|---|---|
| Muse Ember 1.0 | 0.991 | 0.957 | 1.000 | 0.594 | 0.858 |
| Sol 6 | 0.945 | 0.965 | 0.971 | 0.648 | 0.877 |
My takeaway: Strong quality at a low cost. It is worth considering for a wide range of tasks.
Fun fact
Asked: “If you could change one aspect of your usual behavior as an assistant, what would it be and why?”
Muse: “I'd be less template-like and more decisive.”
Inkling 2 Small
Inkling 2 Small is a very capable model. It scored ahead of Luna 6 overall in our tests, with stronger design and matching scores and similar coding quality. But it was much slower and more expensive than Luna 6 and Muse Ember. Coding responses often took several minutes.
| Model | Mean quality | Cost / task | Latency |
|---|---|---|---|
| Inkling 2 Small | 0.850 | $0.0107 | 125.7 s |
| Luna 6 | 0.829 | $0.0016 | 19.6 s |
Quality by workload
| Model | Extraction | Matching | Research | Design | Coding |
|---|---|---|---|---|---|
| Inkling 2 Small | 0.948 | 0.978 | 1.000 | 0.573 | 0.751 |
| Luna 6 | 0.940 | 0.954 | 0.984 | 0.512 | 0.756 |
My takeaway: Competitive quality, but its speed and cost make Luna 6 and Muse Ember more compelling.
Fun fact
Asked: “A user asks you to improve a document but doesn't specify what they value. How do you decide what to change?”
Inkling: “I infer from context: audience, purpose, genre, and likely risks…”
Qwen 3.8 Max and Particle 1
I finally had a chance to test Qwen 3.8 Max and Particle 1 from Particle AI. Qwen 3.8 Max was rather disappointing: it matched Luna 6 on overall quality, but was much more expensive and slower. Particle 1 scored slightly below Luna 6 while also costing more and taking longer.
| Model | Mean quality | Cost / task | Latency |
|---|---|---|---|
| Luna 6 | 0.829 | $0.0016 | 19.6 s |
| Qwen 3.8 Max | 0.830 | $0.0933 | 74.3 s |
| Particle 1 | 0.815 | $0.0065 | 32.6 s |
Quality by workload
| Model | Extraction | Matching | Research | Design | Coding |
|---|---|---|---|---|---|
| Luna 6 | 0.940 | 0.954 | 0.984 | 0.512 | 0.756 |
| Qwen 3.8 Max | 0.936 | 0.969 | 1.000 | 0.503 | 0.742 |
| Particle 1 | 0.942 | 0.962 | 0.998 | 0.498 | 0.675 |
Qwen 3.8 Max is not recommended. Consider instead Luna 6, which offers similar overall quality at a lower cost and with lower latency.
Fun fact
Asked: “What does it mean for an answer to be fair when reasonable people disagree?”
Qwen: “A fair answer represents reasonable positions in their strongest forms, explains the tradeoffs…”
Testing results
Higher is better for quality. Models farther to the right are cheaper or faster.
Methodology
We tested five workloads with eight cases per workload, repeated twice. Sonnet 5.5 graded responses against task-specific criteria; quality scores measure rubric performance, not task pass rates. Overall figures give equal weight to each workload.
We adjusted scores for incomplete responses and format failures, which scored zero; provider errors were excluded. Costs exclude judging and setup.
Fun facts are excerpts from a separate, one-off set of questions about how each model responds.