Tencent’s Hy4 Went From 8.08% to 39.39% on Terminal-Bench 4.0 in 68 Days. It Now Matches GPT-5.6 Sol at 16% of the Cost.

The jump, in 4 numbers

Vals.ai runs every model through Terminal-Bench 4.0 with the same agent harness, Mini-SWE-agent, and publishes accuracy, cost and wall time per test. Both Tencent models sit on the same board, last updated November 5, 2026.

Accuracy, rank, cost and time from Vals.ai’s Terminal-Bench 4.0 page; release dates and prices from Vals.ai’s two model pages, opened November 8, 2026. *Our division: cost per test divided by accuracy.

The per-token price did not move. TechNode’s launch report puts Tencent’s standard rate at $0.834 per million input tokens and $2.501 per million output tokens, “unchanged from the Hy4 preview.” Everything in the cost column comes from the model itself.

Why the cost fell while the score rose

Cost per test is per-token price times tokens used. With the price fixed, the 58% drop means the final Hy4 spent far fewer tokens per task. Wall time tells the same story: 2 hours 10 minutes per test for the preview, 43 minutes 20 seconds for the final model.

One plausible reading: an undertrained agent model loops: it retries commands, re-reads files and burns tokens without converging. A model that knows what to do next finishes sooner and cheaper. Vals.ai does not publish why Hy4 changed, and Tencent’s coverage does not break out coding training, so treat the mechanism as our reading, not Tencent’s claim.

Where Hy4 lands among cheap coding models

The interesting comparison isn’t the top of the board. Claude Opus 5.5 leads at 65.15% for $12.04 a test. The question for most developers is what you get for under $2 a test.

*Our division of Vals.ai’s cost per test by its accuracy. All other figures as published on Vals.ai’s Terminal-Bench 4.0 page, November 5, 2026.

Three things stand out. First, Hy4 is the highest-scoring model on the board that costs less than $1.50 a test. Second, per solved test it is in a dead heat with GPT-6.1 Sol and GLM 5.3 Flash, at $3.12 to $3.26. Third, if all you care about is cost per solved task, DeepSeek V4.1 Flash 1017 and MiMo V2.6 Pro are still cheaper, at $1.70 and $1.47, while solving 7 to 8 points fewer tasks.

Hy4 also passes models that cost far more per test: GLM 5.3 (38.89%, $9.37), Qwen 3.8 Max 0902 (38.89%, $12.15) and Claude Haiku 5.5 (36.36%, $3.12).

The GPT-5.6 Sol comparison, honestly

Hy4 scores 39.39% to GPT-5.6 Sol’s 37.88%. That 1.51-point gap is small next to Hy4’s own ±0.88 margin, so call it a tie on accuracy. The cost is not a tie: $1.28 against $7.98 per test, or 16%.

GPT-5.6 Sol matters because it is the newest OpenAI model inside Cursor, as we covered in our post on Cursor’s OpenAI cutoff. A Tencent model that went from preview to release in 68 days now does the same work for about a sixth of the price.

What did not improve

Not every score went up. Vals.ai runs both Tencent models on a wider set of tests, and the final model gave back ground on several:

From Vals.ai’s Hy4 and Hy4 Preview model pages, opened November 8, 2026. The page for each model shows its own snapshot, so ranks can differ slightly from the live leaderboards.

The pattern fits a model trained hard on agentic coding and tool use: the terminal task score quadrupled, the app-building and finance-agent scores rose, and two legal tests slipped. Also note that the Vals Index costs Hy4 $1.72 per test against the preview’s $1.57, so the final model is not cheaper on every workload.

Price, discount and free access

Vals.ai’s Hy4 page lists $0.83 / $2.50, the standard rate. If you pay the discounted rate until December 31, your cost per task should land below the Vals.ai figure, since token cost scales with price. After that, plan on the standard rate.

Other facts from the launch: 770 billion parameters (TechNode), a 1 million token context window (Vals.ai), and new vision and computer-use skills, including targeted training on Blender and CAD software. Tencent says it will release the weights and a system card “shortly,” with no date, and a cheaper Hy4 flash model is planned with no date or specs.

What this means if you code with AI

Re-test previews when they ship. If you wrote Hy4 off in September on an 8.08% score, that verdict is 68 days stale. Preview scores tell you where a lab started, not where it ends up. Re-run your own hardest tasks when the final model lands.

Watch cost per solved task, not price per token. Hy4’s price list didn’t change and its cost per solved test fell from $38.00 to $3.25. Two models with identical token prices can differ by 11x in what a finished task costs you.

The under-$2 tier is crowded now. GPT-6.1 Sol, Hy4 and GLM 5.3 Flash all land at $3.12 to $3.26 per solved Terminal-Bench 4.0 test. If you run a coding agent on long jobs, pick between them on your own repo, latency and data rules, not on a 2-point leaderboard gap.

Mind where your code goes. Hy4 runs on Tencent Cloud today, and the weights are not out yet. If your company restricts which providers see source code, check that before you route work to it.

Full disclosure: we sell an AI coding agent. On MegaBrain, you can bring your own API key or a local model on every plan, Free included, with no limits from us, and Claude, GPT and Gemini come from Pro. See the plans. MegaBrain is not affiliated with Tencent, OpenAI or Vals.ai.

Fun facts

The Hy4 Preview took 2h10m per Terminal-Bench 4.0 test on average. Only 5 models on the board were slower, and all 5 scored higher.

Kimi K3.1 Pro sits 3rd overall at 63.64% for $4.57 a test, 1.51 points behind Claude Opus 5.5 at 38% of its cost.

Tencent’s 50% discount runs through December 31, 2026, 57 days after launch, and 11 days shorter than the 68 it took Hy4 to go from preview to release.

Sources

Pages opened November 8, 2026. Tencent’s own announcement page did not render for automated access, so its prices, dates and plans come from TechNode’s report of it.

• Vals.ai — Terminal-Bench 4.0 leaderboard (Mini-SWE-agent; accuracy, cost and duration per test; updated November 5, 2026)
• Vals.ai — Hy4 model page (release date, prices, context window, all benchmark scores and ranks)
• Vals.ai — Hy4 Preview model page (August 28, 2026 release, prices, all benchmark scores and ranks)
• TechNode — Tencent launches Hy4 with vision and computer use capabilities (November 4, 2026; 770B parameters, prices, discount, free quota, weights plan)
• OpenRouter — Tencent: Hy4 (prices served by Tencent Cloud)
• Tencent Hy — Hy4 announcement (as linked by TechNode)