New models tested: Fable 5.2, Muse Ember 1.0, Inkling 2 Small

Independent benchmark platform Metergraph evaluated five new artificial intelligence models across standardized production workloads, according to metergraph.dev's report. Evaluator Vasiliy Radostev published performance, pricing, and latency data on 11 November 2026 covering Fable 5.2, Muse Ember 1.0, and Inkling 2 Small, alongside comparative assessments of Qwen 3.8 Max and Particle 1. The tests measured mean output quality across extraction, matching, research, design, and coding rubrics graded by Sonnet 5.5.

This analysis breaks down the benchmark results, unit economics, and latency tradeoffs for engineering teams selecting model backends. The findings reveal diverging development directions, with frontier models posting higher design scores alongside steep cost increases, while smaller specialized models deliver near-frontier quality at a fraction of the operating expenditure.

01Benchmark scores across five workloads

Vasiliy Radostev structured the benchmark around five distinct workloads, evaluating eight individual test cases per workload across two full iterations. Sonnet 5.5 scored each response against workload-specific rubrics to compute a mean quality score from zero to one. Format failures and incomplete responses received zero, while infrastructure provider errors were excluded from totals.

Fable 5.2 achieved the highest overall mean quality at 0.884, driven by a 0.731 design score and a 0.787 coding mark. Muse Ember 1.0 reached a 0.880 mean quality score, logging a 1.000 score in research and 0.991 in extraction. Inkling 2 Small scored 0.850 overall, outperforming Luna 6 in design and matching while securing a 1.000 research score.

Evaluation runs on Qwen 3.8 Max yielded a 0.830 overall quality score, falling below Sol 6 and Muse Ember 1.0. Particle AI's Particle 1 model recorded an overall quality score of 0.815 across the five workloads, scoring 0.498 on design and 0.675 on coding tasks.

02Frontier capabilities versus operating costs

The benchmark demonstrated that capability gains at the highest tier carry significant cost premiums. Fable 5.2 elevated design capability from 0.471 in version 5.1 to 0.731, yet cost per task rose from $0.2833 to $0.4279 while average response latency climbed from 44.6 seconds to 77.2 seconds. Radostev noted that frontier intelligence is becoming more expensive faster than it is improving.

In contrast, Muse Ember 1.0 delivered mean quality comparable to Sol 6 while drastically lowering execution expenses. Muse Ember averaged $0.0053 per task compared to $0.0291 for Sol 6, paired with a 16.1-second latency against Sol 6's 12.2 seconds. Inkling 2 Small logged a $0.0107 cost per task, but averaged 125.7 seconds latency, with complex coding evaluations requiring multiple minutes to complete.

03Workload efficiency and task allocation

The data indicates engineering teams should separate high-reasoning creative workloads from data extraction pipelines. For complex system design and advanced coding architecture, Fable 5.2 establishes the highest performance ceiling among the tested group, outperforming Fable 5.1's 0.684 coding score with its 0.787 mark.

For high-volume production tasks, Muse Ember 1.0 provides an alternative to Sol 6. Its 0.991 extraction score exceeds Sol 6's 0.945, while matching Sol 6 closely on coding with 0.858 versus 0.877, achieving an 81 percent reduction in per-task inference costs.

Inkling 2 Small offers a viable alternative to Luna 6 for quality-focused workloads, logging a 0.573 design score compared to Luna 6's 0.512. However, Luna 6 remains the faster and more economical choice for production pipelines requiring low latency, operating at $0.0016 per task with a 19.6-second turnaround.

04CLI configuration and deployment commands

Developers integrating updated models into local coding agents can update package distributions and switch active model flags directly from the command line.

npm install -g @anthropic-ai/claude-code

To switch the active model during an interactive coding session, execute the model selector command:

/model fable

To initialize a pinned session directly from your terminal shell, pass the target model name with the command-line flag:

claude --model fable

To configure the model as your persistent system default across future interactive sessions, update your agent configuration:

/model default fable

05Latency bottlenecks and production selection

Deploying models requires balancing benchmark quality against response duration. Inkling 2 Small demonstrated competitive baseline quality but logged the highest latency in the benchmark suite at 125.7 seconds. For production services with strict response timeouts, high latencies present significant deployment friction despite solid 0.978 matching and 1.000 research scores.

Testing also highlighted models that failed to justify their price points. Qwen 3.8 Max matched Luna 6's overall score with 0.830 while charging $0.0933 per task and requiring 74.3 seconds to respond. Particle 1 underperformed Luna 6 at 0.815 mean quality while costing $0.0065 with a 32.6-second turnaround.

Radostev advised engineering teams against deploying Qwen 3.8 Max in production pipelines, recommending Luna 6 instead for equivalent output quality at lower cost. For broader workloads, Muse Ember 1.0 stands out as the most balanced alternative for teams seeking near-frontier capabilities without the severe pricing curves of top-tier models.