AI Dev Weekly is a Thursday series where I cover the week’s most important AI developer news, with my take as someone who actually uses these tools daily.

Four stories this week. Anthropic shipped Claude Opus 5 at half Fable 5’s price, and the benchmarks are not close. We compared every major flash model and found a clear winner. AI image and video generation APIs now range from $0.002 to $4.20 per unit, and the gap matters. Plus, the terminal AI coding tool landscape just got clearer.

1. Claude Opus 5: the new frontier at half price

Anthropic released Claude Opus 5 on July 24. The headline: 43.3% on Frontier-Bench, more than double Opus 4.8’s 21.5% and well ahead of Fable 5’s 33.7%. At $5/$25 per million tokens, it costs exactly half of Fable 5.

The numbers:

• $5/$25 per 1M tokens (standard), $10/$50 (fast mode, 2.5x speed)
• 1M context window with 128K max output
• 5-level effort control (min/low/medium/high/max)
• Thinking on by default

The benchmarks:

Benchmark	Opus 5	Fable 5	Opus 4.8
Frontier-Bench v0.1	43.3%	33.7%	21.5%
ARC-AGI 3	3x next best	N/A	N/A
CursorBench 3.2	Within 0.5%	Baseline	N/A

The effort level system is the differentiator. Five thinking levels let you pay for reasoning only when you need it. Min for formatting, max for research problems. Fast mode at $10/$50 matches Fable 5’s price but runs 2.5x faster.

Same-day support on Claude Code, Cursor, OpenRouter, and Aider.

My take: This is the model most developers should be using. At $5/$25, it undercuts GPT-5.6 Sol ($6/$18) while outperforming it on Frontier-Bench. Fable 5 is now only justified for the hardest 10% of problems. See our Opus 5 vs Fable 5 and Opus 5 vs GPT-5.6 Sol comparisons.

2. Flash models compared: Gemini 3.6 Flash leads, but the gap is closing

We published a full comparison of every major flash model: Gemini 3.6 Flash, DeepSeek V4 Flash, Qwen 3.7 Flash, and Step 3.7 Flash. These are the models you use when speed and cost matter more than peak intelligence.

The rankings:

Model	Price (in/out)	Speed	Context	Best For
Gemini 3.6 Flash	$1.50/$7.50	304 tok/s	1M	General purpose, computer use
DeepSeek V4 Flash	$0.28/$1.10	180 tok/s	128K	Budget coding, high volume
Qwen 3.7 Flash	$0.03/$0.15	150 tok/s	1M	Ultra-budget, Chinese models
Step 3.7 Flash	$0.50/$2.00	200 tok/s	256K	Balanced speed/cost

My take: Gemini 3.6 Flash wins on capability (built-in computer use, 1M context) but DeepSeek V4 Flash at $0.28/$1.10 is 5x cheaper. For high-volume API work, DeepSeek is the value king. For interactive coding where latency matters, Gemini’s 304 tok/s is hard to beat. Qwen 3.7 Flash at $0.03/$0.15 is basically free, but you get what you pay for.

3. AI image and video APIs: pricing from $0.002 to $4.20

Two new pricing guides this week cover the image generation and video generation API markets. The range is massive.

Image generation (per image):

• FLUX Schnell: $0.003 (fast, good enough for most use cases)
• FLUX Pro: $0.05 (higher quality, slower)
• DALL-E 3: $0.04-$0.08 (OpenAI, reliable)
• Stable Diffusion XL: $0.002 (cheapest, self-hostable)

Video generation (per second):

• Kling 3.0: $0.09-$0.45 (cheapest, good quality)
• Runway Gen-4.5: $0.50-$2.00 (best quality, most expensive)
• Veo 3.1: $0.15-$0.75 (Google, balanced)

My take: Image generation is now commoditized. At $0.002-$0.08 per image, the cost is negligible for most applications. Video generation is where the real pricing decisions matter. Kling 3.0 at $0.09/sec is 5x cheaper than Runway, and the quality gap is narrowing. If you are building video generation into a product, run your own benchmarks before committing.

4. Terminal AI coding tools: Claude Code vs Aider vs Continue.dev

Our new comparison of the three major terminal AI coding tools breaks down the real differences.

The verdict:

• Claude Code: Best capability, highest cost, locked to Anthropic models
• Aider: Best model flexibility, works with any provider, steeper learning curve
• Continue.dev: Best price (free tier), VS Code integration, least powerful

My take: If you are already paying for Claude, use Claude Code. If you want to use DeepSeek, Qwen, or local models, use Aider. If you want a free option that works in VS Code, use Continue.dev. The choice is simpler than the comparison articles make it seem. Most developers should try Claude Code first, then switch to Aider if they need model flexibility.

Quick hits

• Opus 5 vs DeepSeek V4 Pro: Full comparison of proprietary vs open source frontier models
• Opus 5 vs Kimi K3: Benchmark leader vs open weights challenger
• AI Stack for Solo SaaS Founder: Practical guide on building with AI from coding to support
• AI Context Window Explained: Why 1M tokens is not what you think (spoiler: most of it is marketing)

What I’m watching next week

• Fable 5 price cut? Anthropic cannot ignore the 2x pricing disparity with Opus 5
• OpenAI response. GPT-5.6 Sol is now more expensive and less capable than Opus 5
• Flash model adoption. Which flash model wins the speed/cost/use-case triangle?

---

AI Dev Weekly publishes every Thursday. Subscribe for the newsletter version.

Previous issue: AI Dev Weekly #19