Gemini 3.7 Flash vs GPT-5.6 Terra vs GLM-5.3: Coding
The Short Answer
As of August 2026, the mid-tier coding market has three credible options at three price points. Gemini 3.7 Flash ($0.75/$3.75, released August 13) is the cheapest and posted the largest generational jump. GPT-5.6 Terra ($2/$12) is the most expensive and still edges Gemini on several agentic coding evals. GLM-5.3 ($1.40/$4.40, released August 14) sits in the middle with the strongest open-line long-horizon coding scores. Pick on eval results in your own repo, not on leaderboard averages.
Quick Comparison
| Gemini 3.7 Flash | GPT-5.6 Terra | GLM-5.3 | |
|---|---|---|---|
| Vendor | OpenAI | Z.ai | |
| Released | Aug 13, 2026 | Jul 2026 (price cut Jul 30) | Aug 14, 2026 |
| Price (per MTok) | $0.75 / $3.75 (intro) | $2 / $12 | $1.40 / $4.40 |
| Price after Jan 1, 2027 | $1.50 / $7.50 | $2 / $12 | $1.40 / $4.40 |
| 30K-in/5K-out task | ~$0.041 | ~$0.12 | ~$0.064 |
| Cached input | — | — | $0.26 |
| Context | 1M tokens | Long-context | Long-context |
| Max output | 65,536 | — | — |
| Headline coding score | FrontierCode 1.1: 43.6% · DeepSWE v1.1: 65.3% | Edges Gemini on several agentic evals | Terminal-Bench 3.0: 28.3 |
| Subscription route | Google AI plans | ChatGPT tiers | GLM Coding Plan from $18/mo |
Where Each Wins
Gemini 3.7 Flash — best price-per-capability right now. The generational deltas are the story: FrontierCode 1.1 went 34.4% → 43.6%, AutomationBench 17.0% → 30.4%, GDP.pdf 22.0% → 34.0%. Nearly doubling AutomationBench matters more than the code-gen numbers, because it measures multi-step completion — exactly what an agent loop does. At $0.75/$3.75 with a 1M context, it is the default value pick through 2026.
GPT-5.6 Terra — best reliability in the tier. Terra got a 20% price cut on July 30, 2026 (from $2.50/$15 to $2/$12) and remains the mid-tier model that most consistently finishes hard agentic tasks. Google’s own published comparison concedes Terra the lead on several agentic coding evals. If your failure cost is high but Sol-tier pricing isn’t justified, Terra is the safe middle.
GLM-5.3 — best price stability and subscription math. Z.ai’s August 14 release posts 28.3 on Terminal-Bench 3.0 (up from 4.6 on GLM-5.2) and leads open-line models on CyberGym and AutomationBench. Its $1.40/$4.40 is a standing price, not an intro rate, and cached input drops to $0.26. The GLM Coding Plan from $18/month undercuts every Western subscription. Caveats: thinking is mandatory (inflating output tokens), and open weights are staged rather than available at launch.
The January 1, 2027 Trap
Gemini 3.7 Flash’s price advantage is explicitly dated. Google published the expiry: on January 1, 2027 it goes to $1.50/$7.50 — double the intro rate and above GLM-5.3’s standing price on both sides of the meter.
Cost per reference task, before and after:
| Model | Through Dec 31, 2026 | From Jan 1, 2027 |
|---|---|---|
| Gemini 3.7 Flash | $0.041 | $0.083 |
| GLM-5.3 | $0.064 | $0.064 |
| GPT-5.6 Terra | $0.12 | $0.12 |
If you are architecting a system that runs into 2027, model both columns. A migration that pencils out at $0.041/task and breaks at $0.083/task is not a migration — it is a deferred problem.
The Decision Framework
- High-volume, retry-tolerant loops (test generation, batch migrations, lint fixes) → Gemini 3.7 Flash through 2026, then re-run the math.
- Failure cost dominates token cost (production refactors, unattended agents) → GPT-5.6 Terra, or escalate to a frontier model entirely.
- Long multi-year cost planning + subscription economics → GLM-5.3, whose price isn’t on a countdown.
- Data cannot leave your infrastructure → none of these; use a self-hostable open-weight model.
The honest answer for most teams: these three are close enough that your repo decides. Run 50 real tasks through each, measure success rate and total tokens burned including retries, and let price-per-task rather than price-per-token settle it.
Last verified: August 16, 2026. Prices from vendor pricing pages.