Practical guide. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.
For: Curious readers and builders getting oriented
⚡ TL;DR — the 30-second version
- Start with modality, deployment and policy constraints.
- Evaluate exact IDs rather than product brands.
- Compare total accepted-task cost, not only token rates.
The current reviewed shortlist is Claude Opus 5.5, GPT-6 Astra/Sol/Luna, Gemini 3.8 Flash and Grok 4.7. Treat provider positioning as a hypothesis: Anthropic emphasizes long-running coding and knowledge work, OpenAI differentiates end-to-end reasoning from high-volume work, Google combines multimodal input and agentic workflows, and xAI emphasizes coding and knowledge work.
Eliminate models that cannot meet hard constraints first. Required output modality, region, data-retention policy, available tools, rate limits and license can matter more than a benchmark lead. Gemini Flash's ability to accept audio, for example, does not make it a live speech-output model.
“A shortlist is useful. A permanent winner is not. The right model is the one that clears your acceptance criteria at an acceptable total cost.”
Then run a small, fixed task set representative of your product. Measure accuracy, latency, intervention, refusals and total accepted-task cost. Include caching and long-context billing conditions; do not assume the advertised input rate applies to every request.
This is a source-backed decision framework, not a claim of twelve weeks of testing. The model index links specifications and reported benchmark evidence and leaves missing comparable measurements empty.
Field guide / 01
Turn a preference into a decision matrix
List constraints and give each one a reason. Data residency is a hard constraint when a contract requires it; writing style is usually a preference that can be reviewed against examples. Do not assign a single blended score before separating these categories.
For each candidate, document exact API capabilities, available tools, region, rate limits and the conditions attached to pricing. A lower-cost model that cannot access the required tool is not actually a candidate for that workflow.
Field guide / 02
Keep comparisons small enough to understand
Select a limited task set, change one variable and review each failure. If you change model, prompt, retrieval and tool configuration at once, a better outcome will not tell you which change helped.
Schedule a repeat review when a new release changes a meaningful constraint or when your workload changes. Do not rotate models purely to keep a dashboard looking current.
Before you ship
0/5Use this as a review aid. Checks reset when you leave this page.
Questions that come up
What is the best model overall?
There is no workload-independent answer. Start with the current catalog, remove models that fail hard requirements and evaluate the remaining candidates on your tasks.
Sources & verification
Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.
- 01Claude models overviewAnthropic↗
- 02GPT-6 Astra API model and pricingOpenAI↗
- 03GPT-6 Sol API model and pricingOpenAI↗
- 04GPT-6 Luna API model and pricingOpenAI↗
- 05Gemini 3.8 Flash model specificationsGoogle↗
- 06Grok API — models and pricingxAI / SpaceXAI↗