Workforce
Of The Future.
The daily briefing ↗
THE MODEL DECISION

Match the model
to the work.

A capable model is only useful when its strengths fit your workflow. Compare specific models on the work they suit, the trade-offs they introduce and the control you retain.

8 model assessments·Primary sources reviewed 20 September 2026

How to read this: capabilities are grounded in the linked provider documentation. Pros, cons and fit are our editorial assessment, not results from our own benchmark. Hosted prices, availability and model aliases change; verify them before committing.

OpenAI · General-purpose reasoning model

GPT-6 Astra

Hosted API; model ID gpt-6-astra. No published weights for local deployment.

Best fit: difficult research, coding and document workflows with many dependent steps.

The advantages

  • 1.05M-token context and 128K output accommodate large working sets.
  • Responses API supports search, shell, computer use, MCP and structured outputs.
  • Selectable reasoning effort lets difficult steps receive more computation.

The limitations

  • Premium hosted rates make long, repeated reasoning runs costly.
  • No fine-tuning support; adaptation relies on instructions, retrieval and tools.
  • No native audio or video input; use separate media processing.
OUR ASSESSMENT

A capable central planner with broad tool support. The application still supplies permissions, execution and success checks.

Decision rule: Escalate to Astra when a cheaper model repeatedly fails your real end-to-end tasks and the improvement justifies its cost.

OpenAI · General-purpose reasoning model

GPT-5.6 Terra

Hosted API; model ID gpt-5.6-terra. No published weights for local deployment.

Best fit: everyday tool-using agents and repeated production tasks.

The advantages

  • Supports the same broad Responses tool categories, including MCP and computer use.
  • Reasoning can be disabled or increased, allowing task-specific effort.
  • 1.05M-token context and lower standard token prices than Astra.

The limitations

  • A cost-balanced tier: test difficult planning before assigning unattended work.
  • No fine-tuning support and no local weight deployment.
  • Very large prompts incur higher rates; tool calls add their own costs.
OUR ASSESSMENT

Useful as a default worker in a routed system, with harder or repeatedly failing cases sent to a stronger model.

Decision rule: Start with Terra for a repeatable workload, measure completion cost including retries, then compare Astra on the failing cases.

Anthropic · General-purpose reasoning model

Claude Opus 5

Active hosted model; claude-opus-5 on the Claude API and supported cloud platforms.

Best fit: complex coding, enterprise analysis and iterative project work.

The advantages

  • 1M context and 128K output support substantial code and document work.
  • Adaptive thinking and adjustable effort let the agent spend effort per task.
  • Mid-conversation tool changes can preserve the prompt cache (beta).

The limitations

  • Higher token prices and comparative latency than Sonnet 5.
  • Thinking is enabled by default; disabling it requires effort high or below.
  • Text/image input and text output; separate services are needed for native speech.
OUR ASSESSMENT

Anthropic's recommended general starting point for complex agent work. A Messages API integration still needs its own execution loop.

Decision rule: Use Opus as the Claude baseline; increase effort before paying for Fable, and compare both on task completion.

Anthropic · General-purpose reasoning model

Claude Fable 5.1

Active hosted model, released 1 September 2026; Claude API ID claude-fable-5-1.

Best fit: demanding multistep research and long-running coding or office tasks.

The advantages

  • Designed for long-horizon work across code, documents, spreadsheets and slides.
  • 1M-token context with 128K output.
  • Per-message effort and readable between-tool progress updates are available in beta.

The limitations

  • Higher token prices and slower comparative latency than Opus 5.
  • Adaptive thinking is always on, limiting lightweight execution choices.
  • Forced tool use errors; older models cannot read its thinking blocks, complicating migration.
OUR ASSESSMENT

An escalation model for difficult jobs. Progress updates help humans follow work but do not prove that the result is correct.

Decision rule: Choose Fable when Opus at higher effort still misses important cases in your own evaluation set.

Google · Multimodal general-purpose model

Gemini 3.8 Flash

Stable Gemini API model; ID gemini-3.8-flash. September 2026 update.

Best fit: agents combining documents, images, audio or video with grounded tools.

The advantages

  • Accepts text, images, video, audio and PDFs in one model.
  • Supports search grounding, file search, code execution and function calls.
  • About 1M input tokens plus caching support large multimodal working sets.

The limitations

  • Produces text; image/audio generation and Live API are separate models.
  • Computer use remains a preview capability despite the stable base model.
  • Supports low/medium/high thinking; requesting minimal produces an error.
OUR ASSESSMENT

A practical candidate for agents that must inspect varied media before choosing tools. Media support is a capability, not a measured accuracy ranking.

Decision rule: Shortlist Flash when multimodal input is central, then test grounding accuracy and full-workflow latency on your own files.

DeepSeek · Open-weight multimodal reasoning model

DeepSeek V4.1 Flash

Hosted API ID deepseek-flash; MIT-licensed weights at deepseek-ai/DeepSeek-V4.1-Flash. Released 10 September 2026.

Best fit: input-heavy coding agents and teams evaluating self-hosted infrastructure.

The advantages

  • Native image/text input, tool calls and a 1M context window.
  • Supports Responses and Anthropic-compatible APIs, easing harness integration.
  • Published weights and inference/evaluation resources enable deployment control.

The limitations

  • Hundreds of billions of total parameters make local hosting a substantial infrastructure task.
  • deepseek-flash is a moving service name; legacy V4 names now route to V4.1.
  • API token prices do not represent the hardware, serving and operations cost of self-hosting.
OUR ASSESSMENT

Works within an agent harness; downloadable weights do not supply tool execution or a complete production system.

Decision rule: Compare hosted total task cost first; consider self-hosting when control and workload volume justify the infrastructure.

Qwen / Alibaba · Open-weight multimodal model

Qwen3.8-Flash-Next

Released weights at Qwen/Qwen3.8-Flash-Next under Qwen Community 1.0; experimental architecture preview.

Best fit: teams building custom multimodal and coding agents with control of serving.

The advantages

  • Published weights support self-hosting and inspection of the model configuration.
  • Vision encoder plus tool-oriented examples support multimodal agent work.
  • Hybrid sparse attention targets efficient long-context inference; native context is 262,144 tokens.

The limitations

  • A new experimental architecture demands compatible, current inference software.
  • 6B active parameters is not its memory footprint: large backbone and embedding weights must still be stored.
  • Qwen Community 1.0 is a model-specific licence; do not assume Apache terms or hosted-service features.
OUR ASSESSMENT

Promising for custom infrastructure. Qwen3.8-Flash, the managed service, adds production features and is not identical to the weight package.

Decision rule: Choose it when deployment control matters and your team can validate serving, licence terms and tool reliability.

Mistral AI · Open-weight hybrid reasoning model

Mistral Small 4

Generally available; API ID mistral-small-2603 and Apache 2.0 weights. Released 16 March 2026.

Best fit: configurable business agents where open weights and a permissive licence matter.

The advantages

  • Combines instruction-following, reasoning and coding in one model.
  • Function calling and structured outputs support controlled application integration.
  • Apache 2.0 weights provide deployment flexibility alongside a hosted API.

The limitations

  • 256K context is smaller than the million-token models in this comparison.
  • 119B total parameters still require substantial memory despite only 6.5B being active.
  • Hosted built-in tools and agent endpoints are services; self-hosting weights does not include those services.
OUR ASSESSMENT

A candidate for repeatable, tool-driven workloads. Mistral's agent service is a separate layer around the model.

Decision rule: Shortlist Small 4 when licensing and hosting choice matter, then test your longest inputs and hardest tool sequences.

MAKE THE CHOICE WITH EVIDENCE

Run the same job. Count the whole cost.

Quality

Use representative cases with clear acceptance criteria. Include ambiguity, tool failures and incomplete data.

Economics

Measure cost per accepted result, including retries, tools and human review. Token price alone misses the work.

Control

Check data handling, hosting options, permissions, version stability and a practical exit path.

Operations

Measure latency at busy times, recovery and escalation. Use the same harness and settings for a fair comparison.

A DIFFERENT KIND OF MODEL

Not every decision needs a conversation.

Jev focuses on bounded, typed decisions. Explore where that fits alongside general-purpose models.

Read about Jev
FIND YOUR NEXT IDEA

Explore Workforce Of The Future.

Articles, AI models, innovations and the worldwide robot directory.