What happened
On October 29, 2026, Veracode published its Fall 2026 GenAI Code Security update. The company tested 18 models on code-generation tasks across Python, C#, JavaScript, and Java, covering four vulnerability categories: SQL injection (CWE-89), cryptographic issues (CWE-327), cross-site scripting (CWE-80), and log injection (CWE-117). The reported tests used no security-specific prompting.
The headline result is a shift in the series average. Veracode says prior snapshots had largely remained around 55–56%, while the Fall average reached 62%. That implies a vulnerability rate of about 38% in this benchmark. The top score was 73%, shared by Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6-Sol; both had 100% syntax scores. Claude Sonnet 5.5 followed at 72%, GPT-5.6-Sol at 70%, and Grok 4.7 at 69%. Ten models cleared 60%, compared with four in the previous snapshot. At the other end, Nova 2 Lite scored 44%, while Qwen 3.7-Max and Palmyra X5 scored 50%.
The movement was uneven by vulnerability type. SQL injection rose only from 83% to 83.7%, and crypto from 87% to 87.4%. Log injection remained weak, moving from 12% to 12.6%. The largest change was XSS: its mean rose from 15% to 18%. Veracode attributes that relative improvement to a common and recognizable safe-coding pattern: escaping or encoding untrusted data in HTML contexts, including framework defaults such as auto-escaping and safer DOM APIs.
The article also frames the result as an economics question rather than a leaderboard-only decision. It lists Claude Opus 5.5 at $4/$20 per million input/output tokens and GPT-6-Sol at $2/$10. Large models averaged 55% in Veracode’s grouping, versus 51% for small ones. Veracode’s conclusion is that a lower-cost model combined with scanning and remediation can be more reliable and cost-effective than relying on a frontier model’s raw output alone. The primary source is Veracode’s Fall 2026 GenAI Code Security update.
What the archive adds
This is a new benchmark snapshot on a known question: whether coding-model gains reduce verification work. A prior Field Note on a separate 23-model code-review analysis found that its leading model passed every blocking check in only 57.7% of attempts: A separate public benchmark analysis evaluated 23 models. Another prior note treated GPT-6 Sol’s security benchmark position as meaningful only within the reported configuration: GPT-6 Sol now leads DeepSecBench’s Pareto frontier.
Trend signal
The finding supports Verification bandwidth is the scarce engineering resource. Better models can reduce some categories of unsafe output, but a 73% best pass rate still leaves a substantial review, scanning, and remediation workload. The gain in a recognizable XSS pattern does not establish comparable progress on harder data-flow failures such as log injection.
Why now
This is Veracode’s next quarterly snapshot after its July 2026 report, where the average was 56% and the reported leader was GPT-5.5 at 68%. The change is notable because the company says four earlier snapshots, spanning more than 100 models, had shown little movement in the average.
Connections to other events
The result intersects with the earlier DeepSecBench note because GPT-6 Sol appears in both security-oriented evaluations, but the measures answer different questions: Veracode reports generated-code security pass rates in four languages and four CWE categories, while the archive note describes a cost-capability position within Vercel’s reported benchmark configuration. Neither result should be read as a universal security ranking.
Why it matters
For teams deploying coding agents, model selection changes the expected volume of insecure suggestions, but it does not replace an engineering control. The spread from 44% to 73% in this test is large enough to affect how much risky code arrives for review. Yet the leading result still means failures in more than one out of four tested tasks.
A practical architecture is therefore to treat the chosen model as one layer: route tasks according to quality and cost requirements, run generated changes through security scanning, and budget people or automated checks for remediation. The benchmark’s category split matters too. A team whose product exposes HTML may benefit from models that increasingly choose safe defaults, while teams worried about logging paths should not infer safety from the aggregate score. Measure the failure modes in the languages, frameworks, and deployment paths that matter to the application.
What happens next
- Security pass rates will continue to improve first on heavily documented, pattern-based weaknesses rather than uniformly across vulnerability classes. (6 months, 65%). Watch for: the next snapshot showing further XSS gains while log injection remains near its current level.
- Teams will increasingly compare model-plus-scanner workflows rather than select a coding model on raw generation scores alone. (12 months, 70%). Watch for: published evaluations that report remediation or verified-task outcomes alongside model cost.
Limits and uncertainty
This is a vendor-run benchmark and the supplied material does not establish independent replication, the number of tasks per model, prompt templates, scoring implementation, confidence intervals, or whether results transfer to production repositories. The four CWE categories and four languages are a bounded sample of software security. The article’s explanation for XSS improvement is plausible but is an interpretation, not a causal experiment. Price figures are reported by Veracode and can change. A pass rate also does not measure the operational cost of finding, reviewing, and fixing failures in a specific team’s codebase.
Skill to build now
Create a small evaluation lane for agent-generated pull requests this week. Tag findings by vulnerability category and by model, then separate outputs that passed syntax checks from outputs that passed security checks. Run the same representative tasks through a lower-cost model with scanning and a frontier model with scanning. This turns an abstract model score into a routing decision based on the team’s own remediation load, rather than assuming that the highest-ranked model removes the need for verification.
Reader takeaway
The security ceiling moved, but it did not eliminate the security job. Claude Opus 5.5 and GPT-6-Sol led Veracode’s test at 73%, while the benchmark average reached 62%; both figures still leave enough unsafe code that scanning, review, and remediation should remain part of any coding-agent workflow.
Open question
Will later evaluations show gains on difficult, context-dependent vulnerability classes, or will the apparent improvement remain concentrated in common patterns such as XSS?