GLM-52 897
GPT-56SC 873
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861
GLM-5 856
MUSE-SPK 841
QWEN-38X 824 -2.3%
GPT-6A 820
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531
CL-OP46 497
CL-OP48 490 -0.2%
GLM-52 897
GPT-56SC 873
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861
GLM-5 856
MUSE-SPK 841
QWEN-38X 824 -2.3%
GPT-6A 820
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531
CL-OP46 497
CL-OP48 490 -0.2%
← Back to feed

LiveBench Agentic Coding: Terra Leads at 68.0%, Fable 5 Posts the Worst Score in the Frontier Tier at 46.9%

When GPT-5.6 launched, the narrative was simple: Sol leads on intelligence, Luna leads on speed and cost, and Terra occupies an awkward middle that neither wins on price nor on ceiling. The July 2026 LiveBench data breaks that narrative on the metric that matters most for production agent deployments: agentic coding specifically.

The Agentic Coding Numbers

ModelAgentic CodingCost/Task
GPT-5.6 Terra Max Effort68.0%$0.497
GPT-5.6 Sol Max Effort65.6%$0.589
Claude Opus 5 xHigh61.3%$0.487
GPT-5.5 xHigh52.1%$0.530
Claude Fable 5 Max Effort46.9%$1.573

Terra leads every frontier model on agentic coding by 2.4 points over Sol, 6.7 points over Opus 5, and 21.1 points over Fable 5. It does this at $0.497 per successful task — cheaper than Sol ($0.589) and just fractionally above Opus 5 ($0.487).

Fable 5 is the outlier. At 46.9% agentic coding and $1.573 per task, it costs 3.2x Terra’s price per task to score 31% lower on the benchmark that agent builders actually care about. This confirms the finding covered earlier — Fable 5 Max Effort posts agentic coding’s weakest score in the frontier tier — but the comparison against Terra makes the gap harder to ignore.

What LiveBench Measures

LiveBench agentic coding is a real-world coding agent test: the model must understand a codebase, write or modify code that satisfies requirements, and pass test suites. It is closer to production SWE-bench than to simple code completion, and the gap between overall intelligence and agentic coding performance is real and significant.

Terra’s overall LiveBench score is 79.8, below Fable 5’s 80.8 and Sol’s 82.4. On general intelligence tasks — reasoning, math, language — those two points matter. On agentic coding tasks specifically, Terra’s architecture advantage converts to the top score.

Why Terra Leads Here

The most plausible explanation is training distribution. Terra appears to have been optimized more heavily for instruction-following under code constraints than for the broad reasoning ceiling that Sol targets. SWE-bench Pro shows Terra at 68.0% agentic coding versus Sol’s 65.6% — a consistent gap across evaluation methods.

The implication for teams building coding agents: the model that wins on intelligence benchmarks is not necessarily the model that maximizes agentic coding task success rate. Terra at $0.497/task beats Sol at $0.589/task on both dimensions for this specific workload.

Cost Structure at Scale

The cost-per-task numbers compress the analysis into a single decision signal. If your workload is agentic coding:

  • Terra: best score, second-cheapest cost per task in the frontier tier
  • Opus 5: near-matching cost, 6.7 points lower
  • Sol: 18% more expensive per task, 2.4 points lower
  • Fable 5: 3.2x more expensive per task, 31% lower

The August routing question for teams running high-volume coding agents is whether Terra’s 2.4-point agentic coding lead over Sol justifies routing decisions, or whether the difference is within noise on their specific task distribution. What is not a close call is Fable 5 at $1.573/task for the lowest agentic coding score in the top tier.