JieGou supports models from Anthropic, OpenAI, and Google. We built it this way because no single model is best at everything — and after running our Recipe Factory pipeline across thousands of automated test executions, the data backs this up.
Here’s what we’ve observed about model performance across real business tasks, not synthetic benchmarks.
Content generation: Claude leads on structure
For tasks like blog post outlines, email drafting, proposal summaries, and customer communications, Claude models consistently produce better-structured output. The writing is organized into clear sections, follows the requested format closely, and maintains a professional tone without being stiff.
Claude Sonnet 4.5 is the sweet spot for most content generation. It’s fast enough for interactive use, produces high-quality prose, and follows output schemas reliably. Opus 4.5 produces marginally better output for complex writing tasks but at significantly higher cost and latency.
GPT-5.1 is competitive on content generation, particularly for shorter outputs like email subject lines, social media posts, and ad copy. It’s strong at matching specific tones and styles when given examples.
Gemini 2.5 Pro handles content generation adequately but tends toward more verbose output. It works well when you want comprehensive coverage of a topic but requires more schema discipline to keep output focused.
Data extraction: Cheaper models are fine
Extracting structured data from unstructured text — invoice processing, resume screening, ticket triage — doesn’t need frontier models. The task is well-defined: read the input, identify the relevant fields, fill in the schema.
Claude Haiku 4.5 and GPT-5-mini both perform well on extraction tasks at a fraction of the cost. They follow output schemas reliably and handle format variations in input text without issues.
Gemini 2.5 Flash Lite is the most cost-effective option for high-volume extraction. Performance is comparable to the other lightweight models at lower token prices.
The key insight: don’t pay for reasoning capability when the task is pattern matching. A model that costs $0.25 per million tokens extracts invoice data just as well as one that costs $15 per million tokens.
Complex analysis: Reasoning models earn their cost
SWOT analyses, contract clause review, deal risk assessment, and strategic planning require the model to consider multiple factors, weigh trade-offs, and produce nuanced conclusions. This is where frontier and reasoning models differentiate.
Claude Opus 4.5 with extended thinking produces the most thorough analyses. The thinking budget (10K tokens) gives it room to work through complex reasoning before producing the final output. It catches edge cases and qualifications that faster models miss.
o3 (OpenAI’s reasoning model) takes a different approach — it uses chain-of-thought reasoning with medium effort by default. The output is strong on logical analysis and quantitative reasoning. It’s particularly good at tasks with clear criteria (deal scoring, compliance checking).
Gemini 3 Pro with reasoning support produces solid analyses but occasionally includes tangential observations that need schema discipline to constrain.
Schema compliance: All modern models are good
One concern teams have is whether the AI will actually follow the output schema. In our testing across thousands of runs, all current-generation models produce valid structured output at rates above 95%. The key factor isn’t the model — it’s the schema definition.
Clear schemas with field descriptions, enum constraints, and examples produce better compliance than minimal schemas that leave the model guessing. A field defined as risk_level (enum: high, medium, low) — Overall risk assessment based on clause analysis gets filled correctly more reliably than risk_level (string).
Web search: Varies by provider
For recipes that need current information — prospect research, competitive analysis, regulatory updates — web search capability matters.
All three providers support web search, but the implementation differs:
- Claude with web search produces well-sourced research with specific citations
- GPT-5.x with web search is strong at synthesizing multiple sources into a coherent narrative
- Gemini with web search benefits from Google’s search infrastructure and tends to surface more diverse sources
For prospect research specifically, we’ve found Claude and GPT produce the most actionable output. For broader market research, Gemini’s search breadth can surface sources the others miss.
The practical recommendation
Most teams don’t need to run benchmarks. Here’s the starting configuration that works for the majority of use cases:
| Task type | Recommended model | Why |
|---|---|---|
| Content generation | Claude Sonnet 4.5 | Best structure and tone |
| Data extraction | Claude Haiku 4.5 | Fast, cheap, accurate |
| Complex analysis | Claude Opus 4.5 | Deepest reasoning |
| Quick classification | GPT-5-mini | Lowest latency |
| High-volume batch | Gemini 2.5 Flash Lite | Lowest cost |
| Research with web search | Claude Sonnet 4.5 | Best-sourced output |
Then optimize from there. Run the same recipe with different models using the same inputs and compare the output quality. JieGou tracks execution time, token counts, and lets you attach quality feedback to each run, making comparison straightforward.
Per-step optimization in workflows
The real power is combining models within a single workflow. A five-step workflow might use three different models:
- Extract data (Haiku) — fast, cheap
- Analyze patterns (Sonnet) — balanced
- Draft summary (Haiku) — fast, cheap
- Generate strategic recommendations (Opus) — highest quality
- Format for email (Haiku) — fast, cheap
Steps 1, 3, and 5 don’t need expensive reasoning. Steps 2 and 4 do. Mixing models at the step level optimizes both cost and quality across the workflow.