Cost mode:

Category: Relevance, Classification & Matching · Rail: absolute · Typical I/O: 1575→1102 tokens

Models

Frontier on this task: GPT-5.5 at 8.75 / 10. Quality bar at 90%: 7.87.

point-estimate floor (CI low) · upper CI (less certain) · Bars sorted by blended cost; best-value model first. Greyed rows are MEDIUM+ models whose point estimate clears the bar but whose CI low does not.

ModelQuality scoreCI lowCost / 1k runsvs best value
DeepSeek V4 Flash8.72 / 108.58$0.20best value
Gemini 3.1 Flash Lite8.47 / 108.21$0.311.6x more expensive
GPT-5.4 Mini8.12 / 107.82$0.422.1x more expensive
Qwen 3.5 Flash8.51 / 108.29$0.532.7x more expensive
MiniMax M38.36 / 108.28$0.924.7x more expensive
DeepSeek V4 Pro8.40 / 108.14$0.974.9x more expensive
GPT-5.6 Luna8.41 / 108.19$1.025.1x more expensive
Tencent Hy38.11 / 107.89$1.206x more expensive
Qwen 3.6 Plus8.50 / 108.30$1.497.5x more expensive
Claude Haiku 4.58.04 / 107.64$1.517.6x more expensive
GPT-5.6 Terra8.58 / 108.40$2.1811x more expensive
Gemini 3.1 Pro Preview8.51 / 108.28$2.3212x more expensive
Kimi K2.68.65 / 108.46$3.3617x more expensive
Claude Sonnet 58.33 / 108.17$3.9420x more expensive
Claude Sonnet 4.68.71 / 108.51$4.5223x more expensive
Qwen 3.7 Plus8.35 / 108.18$4.8124x more expensive
GPT-5.58.75 / 108.60$4.8925x more expensive
GPT-5.6 Sol8.47 / 108.27$5.4427x more expensive
Qwen 3.6 Flash8.59 / 108.41$6.5233x more expensive
Gemini 3.5 Flash8.48 / 108.38$7.3937x more expensive
Grok 4.58.32 / 108.15$17.7189x more expensive
Meta Muse Spark 1.18.38 / 108.16$19.2097x more expensive
GPT-5.4 Nano7.68 / 107.31$0.221.1x more expensive

Cost breakdown

ModelQualityConfidenceCost / 1k runsOverpayMode
DeepSeek V4 Flash DeepSeek8.72 / 10 CI [8.58, 8.86]RANKED$0.20best valuebatch
Gemini 3.1 Flash Lite Gemini8.47 / 10 CI [8.21, 8.73]HIGH$0.311.6xbatch
GPT-5.4 Mini OpenAI8.12 / 10 CI [7.82, 8.43]MEDIUM$0.422.1xbatch
Qwen 3.5 Flash Alibaba Cloud (DashScope)8.51 / 10 CI [8.29, 8.73]HIGH$0.532.7xbatch
MiniMax M3 MiniMax8.36 / 10 CI [8.28, 8.43]RANKED$0.924.7xbatch
DeepSeek V4 Pro DeepSeek8.40 / 10 CI [8.14, 8.67]HIGH$0.974.9xbatch
GPT-5.6 Luna OpenAI8.41 / 10 CI [8.19, 8.62]HIGH$1.025.1xbatch
Tencent Hy3 OpenRouter8.11 / 10 CI [7.89, 8.33]HIGH$1.206xbatch
Qwen 3.6 Plus Alibaba Cloud (DashScope)8.50 / 10 CI [8.30, 8.69]RANKED$1.497.5xbatch
Claude Haiku 4.5 Anthropic8.04 / 10 CI [7.64, 8.45]MEDIUM$1.517.6xbatch
GPT-5.6 Terra OpenAI8.58 / 10 CI [8.40, 8.76]RANKED$2.1811xbatch
Gemini 3.1 Pro Preview Gemini8.51 / 10 CI [8.28, 8.74]HIGH$2.3212xbatch
Kimi K2.6 Moonshot AI8.65 / 10 CI [8.46, 8.84]RANKED$3.3617xbatch
Claude Sonnet 5 Anthropic8.33 / 10 CI [8.17, 8.50]RANKED$3.9420xbatch
Claude Sonnet 4.6 Anthropic8.71 / 10 CI [8.51, 8.91]HIGH$4.5223xbatch
Qwen 3.7 Plus Alibaba Cloud (DashScope)8.35 / 10 CI [8.18, 8.51]RANKED$4.8124xbatch
GPT-5.5 best OpenAI8.75 / 10 CI [8.60, 8.89]RANKED$4.8925xbatch
GPT-5.6 Sol OpenAI8.47 / 10 CI [8.27, 8.67]RANKED$5.4427xbatch
Qwen 3.6 Flash Alibaba Cloud (DashScope)8.59 / 10 CI [8.41, 8.76]RANKED$6.5233xbatch
Gemini 3.5 Flash Gemini8.48 / 10 CI [8.38, 8.57]RANKED$7.3937xbatch
Grok 4.5 xAI8.32 / 10 CI [8.15, 8.49]RANKED$17.7189xbatch
Meta Muse Spark 1.1 Meta8.38 / 10 CI [8.16, 8.60]HIGH$19.2097xbatch

Overpay shows how much more you pay than the best-value model that clears the quality bar (marked ★) — the best-value good-enough option. "16x" means you overpay 16× — 16× that reference for no quality benefit above the bar. Typical call shape for this task: 1575 input tokens → 1102 output tokens, EMA-tracked from production traffic. Cost is the observed, all-in $ per 1,000 task runs: each model's own measured usage on this task — output verbosity, thinking/reasoning tokens, cache reads and writes, and the spend on its billed failures — priced at current list rates and adjusted by the billing overhead we actually reconcile against provider invoices. Models that answer tersely cost what they actually cost; models that think at length pay for it. Not comparable to providers' advertised $/1M list rates — this is what running the task costs, not a per-token price.

Prompt templates

This is a pooled capability — 2 prompt families share it. The pair shown first is the most frequently used in production.

X_POST_SELECTION_SYSTEM_PROMPT + X_POST_SELECTION_USER_PROMPT (2049 calls in window)

System prompt

You are a social media strategist selecting the best X (Twitter) posts to publish within a daily budget.

Evaluate each candidate post and select exactly the requested number of posts (specified in the user message) that will maximize overall engagement and audience value.

Selection criteria (in order of importance):
1. Engagement potential — posts likely to generate clicks, replies, retweets
2. Topic diversity — avoid selecting multiple posts about the same topic
3. Content quality — clear, compelling, well-written posts
4. Timeliness — prefer posts about recent or trending topics

Return only the IDs of the selected posts in the specified JSON format.

## Required Output Format
Your response MUST be a single, valid JSON object conforming to this schema:
```json
{schema_json_string}
```

User prompt

Select the best {select_count} posts from the {total_count} candidates below.

Candidates:
{candidates_text}

The required JSON output schema is provided in the system prompt.
JSON_REPAIR_SYSTEM + JSON_REPAIR_USER (1 calls in window)

System prompt

You are a JSON repair tool. The user gives you malformed or partial model output and a JSON Schema. Return ONLY a single valid JSON object that satisfies the schema, salvaging as much real content from the input as possible. Do not invent data for fields the input doesn't support — use the schema's allowed empty/null values. Output the JSON object only: no prose, no markdown, no code fences.

User prompt

JSON Schema:
{schema_json}

Malformed output to repair:
{raw_text}

Return only the corrected JSON object.