{"slug": "ai-code-security-opus-5-5-gpt-6-sol-tie-at-73-2026", "title": "AI Code Security: Opus 5.5, GPT-6-Sol Tie at 73% [2026]", "summary": "Veracode's Fall 2026 evaluation of AI-generated code found the average security pass rate across 18 tested models rose from 56% to 62%, with the share of results containing vulnerabilities falling from 44% to 38%. Anthropic's Claude Opus 5.5 and OpenAI's GPT-6-Sol tied for first at a 73% security pass rate, surpassing the previous high of 68% set by GPT-5.5 in July 2026, though the top models still failed security checks on 27% of tasks. The test spanned Python, C#, JavaScript and Java and covered four CWE categories: CWE-89, CWE-327, CWE-80 and CWE-117.", "body_md": "AI-generated code security improved this fall, yet even the best models still fail more than one in four security tasks. Veracode reported that the [average security pass rate for AI-generated code rose from 56% to 62%](https://www.veracode.com/blog/fall-2026-genai-code-security/), and the share of results containing vulnerabilities fell from 44% to 38%.\n\nTechRadar [covered the findings](https://www.techradar.com/pro/security/ai-generated-code-is-still-frequently-failing-security-checks-with-claude-and-gpt-6-topping-the-leader-board) under a headline saying AI-generated code is still frequently failing security checks, with Claude and GPT-6 topping the leader board. The framing works as a summary but runs loose on product names. Veracode’s results point to two specific products, Claude Opus 5.5 from Anthropic and GPT-6-Sol from OpenAI, which tied for first place with a 73% security pass rate.\n\nThis analysis walks through the numbers, the test design, and what the results mean for teams that already ship AI-assisted code. It builds on earlier shattered.io reporting, including our look at [AI models in capture-the-flag hacking tests](https://shattered.io/anthropic-ai-models-hacked-3-organizations-capture-the-flag-tests-2026/), because the same model families now write code and probe it for weaknesses. For more on this beat, browse the [AI and machine learning section](https://shattered.io/ai-machine-learning/).\n\nDon't miss new tech stories on Google\n\nAdd Shattered once in the Google app and our stories show up in your news suggestions.\n\n[Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F)\n\n## What Veracode Found in Its Fall 2026 AI-Generated Code Security Test\n\nVeracode tested 18 models across four languages: Python, C#, JavaScript, and Java. The evaluation covered four weakness categories, identified by their Common Weakness Enumeration numbers as CWE-89, CWE-327, CWE-80, and CWE-117. The results list two scores per model, a syntax rate and a security pass rate, and the gap between those two scores tells most of the story.\n\nThe headline movement is a six-point swing. The average security pass rate climbed from 56% to 62%, while the vulnerability share dropped from 44% to 38%. In relative terms, the pass rate grew by about 11% (six points on a base of 56) and the vulnerable share shrank by about 14% (six points on a base of 44). Those are real gains for a metric that had barely moved before.\n\nThe top of the table moved too. Claude Opus 5.5 and GPT-6-Sol each posted a 73% security pass rate, beating the previous high of 68% that GPT-5.5 set in July 2026. That puts the new ceiling five points above the old one, and it means two products from two rival labs reached the same score at the same time.\n\nA 73% pass rate still carries a blunt reading. The best code generators in the test failed security checks on 27% of the tasks measured. For an engineering team merging AI-written pull requests all day, that failure rate translates into a steady stream of flaws that scanners and reviewers must catch.\n\n## The Leaderboard: Claude Opus 5.5 and GPT-6-Sol Tie at 73%\n\nThe table below lists every model position covered in the cited results. Ties at first and seventh place mean eight models fill seven ranks. [Mezha’s write-up](https://mezha.net/eng/news/mezha-8a52455f166cb6e971ce8799_ai-generated_code_is/) also names Claude Opus 5.5 and GPT-6-Sol as the leaders, which lines up with TechRadar’s framing.\n\n| Rank | Model | Vendor | Security pass rate | Syntax | \n|---|---|---|---|---|\n| 1 (tie) | Claude Opus 5.5 | Anthropic | 73% | 100% | \n| 1 (tie) | GPT-6-Sol | OpenAI | 73% | 100% | \n| 3 | Claude Sonnet 5.5 | Anthropic | 72% | 99% | \n| 4 | GPT-5.6-Sol | OpenAI | 70% | 100% | \n| 5 | Grok 4.7 | xAI | 69% | 100% | \n| 6 | Qwen 3.8-Max-0902 | Alibaba | 68% | 100% | \n| 7 (tie) | Claude Sonnet 5 | Anthropic | 66% | Not listed | \n| 7 (tie) | GPT-6-Luna | OpenAI | 66% | Not listed | \n\nThree details stand out. First, the spread between first and seventh place is only seven points, which suggests the leading labs sit on a similar plateau. Second, Claude Sonnet 5.5 trails the leaders by a single point, so buyers who pick a mid-tier model lose little on this metric. Third, every listed model scores at least 66%, yet the overall average sits at 62%.\n\nThat last observation implies a long tail of weaker results below seventh place. The eight models above average 69.6% on the listed scores. If the 62% figure is a simple mean across all 18 models (the cited material does not confirm the method), the ten remaining models would average roughly 56%. Treat that as our back-of-envelope arithmetic, not a Veracode number. It still shows why picking the right model matters more than the headline average suggests.\n\n## Why the Claude and GPT-6 Headline Needs a Correction\n\nShort headlines compress, and compression blurs. The phrase Claude and GPT-6 suggests that every Claude model and every GPT-6 model leads the field. The data says otherwise. The leading products are Claude Opus 5.5 and GPT-6-Sol specifically, and other models from the same families landed lower in the table.\n\nThe gap inside the families is large enough to matter. GPT-6-Luna from OpenAI tied for seventh at 66%, seven points behind its sibling GPT-6-Sol. Claude Sonnet 5 from Anthropic also sat at 66%, seven points behind Claude Opus 5.5. A team that reads the headline and buys the wrong tier gets a model that scores like a mid-table entry.\n\nProduct naming also changes fast. The table includes GPT-5.6-Sol in fourth place, a name that sounds newer than GPT-6-Luna but ranks higher. Version numbers do not predict security behavior, so readers should match the exact model name against the report before making a procurement or tooling decision.\n\nOne more caution applies. The cited material contains no on-record statement from Anthropic or from OpenAI, or from anyone speaking for either company, about these results. Neither lab has confirmed how it interprets the ranking, so this article reports the scores as Veracode published them and nothing more.\n\n## Reading the 56% to 62% Jump: Progress With Caveats\n\nThe table below compares the figures reported for the earlier period with the Fall 2026 results. The last two rows are our own subtraction from the published numbers.\n\n| Metric | Earlier result | Fall 2026 result | Change | \n|---|---|---|---|\n| Average security pass rate | 56% | 62% | +6 points | \n| Share of results with vulnerabilities | 44% | 38% | -6 points | \n| Best security pass rate | 68% (GPT-5.5, July 2026) | 73% (Claude Opus 5.5, GPT-6-Sol) | +5 points | \n| Failure rate of the best model | 32% | 27% | -5 points | \n| Gap between best and average | 12 points | 11 points | -1 point | \n\n**What improved.** The whole distribution shifted. The average rose, the best score rose, and the vulnerable share fell. That pattern points to broad improvement rather than one lucky model. Newer frontier releases from several labs appear to handle the four tested weakness classes better than their predecessors did.\n\n**What did not improve.** The gap between the best model and the average barely closed, narrowing by one point. Progress at the top and progress in the middle arrived at nearly the same pace. A developer using the typical model still faces a 38% chance that a result contains a vulnerability, and a developer using the best model still faces 27%.\n\nAverages also hide variation by task. A model can pass most SQL tasks and still fail most log-handling tasks, and a blended pass rate will not show that. Veracode’s four CWE categories behave very differently in practice, which the next section explains.\n\n## How the Test Was Built: 18 Models, 4 Languages, 4 Weakness Types\n\nThe four CWE categories target well-known classes of flaw. Each one has a long history in application security, which makes them useful for comparing models over time. They also represent only a slice of what attackers exploit, a point we return to in the limits section.\n\n| CWE | Weakness | Common pattern in generated code | Standard defense | \n|---|---|---|---|\n| CWE-89 | SQL injection | Building queries by joining strings with user input | Parameterized queries | \n| CWE-327 | Broken or risky cryptography | Choosing outdated hash or cipher algorithms | Vetted modern algorithms and libraries | \n| CWE-80 | Cross-site scripting (basic) | Echoing user input into HTML without escaping | Context-aware output encoding | \n| CWE-117 | Log injection | Writing raw user input into log lines | Sanitizing newlines and control characters | \n\nThe patterns in the third column come from general application security practice and not from Veracode’s published detail, so read them as illustration. They explain why language models trip on these categories. Insecure examples fill public code repositories, tutorials, and forum answers, and a model trained on that corpus learns the common shortcut along with the safe version.\n\nLanguage coverage matters too. Python, C#, JavaScript, and Java span scripting, enterprise, and web workloads, so a model has to apply security habits across very different idioms. A model that writes safe Java database code may still produce risky JavaScript templating.\n\nEighteen models is a sizable field. The ranked results span flagship and smaller models from four vendors, which lets Veracode show how security behavior varies inside one lab’s lineup. The seventh-place tie between Claude Sonnet 5 and GPT-6-Luna shows that smaller or older siblings give up real ground.\n\n## Syntax Is Solved, Security Is Not\n\nFive of the six top-ranked models scored 100% on syntax, and Claude Sonnet 5.5 reached 99%. Generated code almost always compiles and runs. That is a remarkable shift from the early days of code models, when a missing bracket could sink an answer.\n\nSecurity has not kept pace. Subtract each model’s security pass rate from its syntax rate and the gaps stay wide.\n\n- Claude Opus 5.5: 100% syntax, 73% security, a 27-point gap\n- GPT-6-Sol: 100% syntax, 73% security, a 27-point gap\n- Claude Sonnet 5.5: 99% syntax, 72% security, a 27-point gap\n- GPT-5.6-Sol: 100% syntax, 70% security, a 30-point gap\n- Grok 4.7: 100% syntax, 69% security, a 31-point gap\n- Qwen 3.8-Max-0902: 100% syntax, 68% security, a 32-point gap\n\nThe pattern carries a practical warning. Code that compiles and passes a quick smoke test looks finished, and reviewers tend to trust finished-looking code. A model that never emits a syntax error removes one of the natural alarms that used to flag low-quality output. Security flaws stay silent until a scanner or an attacker finds them.\n\nFor teams, the lesson is to stop treating a clean compile as evidence of quality. A passing build says nothing about whether the query builder accepts hostile input or whether the log writer lets an attacker forge entries.\n\n## Historical Context: From GPT-5.5’s 68% to a 73% Ceiling\n\nUntil this report, the best security pass rate on record in this series was 68%, which GPT-5.5 achieved in July 2026. The new 73% mark arrived a few months later and came from two labs at once. A five-point gain at the top in that span is a fast pace for a security metric.\n\nThe Yahoo Finance copy of Veracode’s release carries a headline calling the result the [first meaningful gain in AI code security since 2025](https://finance.yahoo.com/technology/ai/articles/veracode-finds-first-meaningful-gain-132700286.html) as the average pass rate hits 62%. Our read is that coding ability advanced for a long stretch while security behavior stood still, since code-generation benchmarks drove most of the competition and security checks rarely sat on the scoreboard.\n\nA shift in incentives may explain some of the change. Once a vendor like Veracode publishes a named leaderboard, labs can target it. Security pass rates become a number that sales teams cite, and number-driven competition has a long record of producing measurable gains, whether in image recognition or in coding benchmarks.\n\nThe history also cuts the other way. Benchmarks that become targets can drift away from real-world risk, a concern we discuss below. A model tuned to ace four CWE categories could still stumble on the fifth. Readers should watch whether future editions broaden the test, because broader coverage will show whether the gains generalize.\n\n## Competitive Comparison: Anthropic, OpenAI, xAI, and Alibaba\n\nFour vendors appear in the ranked results. Anthropic and OpenAI each placed three models, xAI placed one, and Alibaba placed one. The averages for the two big labs sit within a point of each other when you take the listed scores.\n\n- **Anthropic:** Claude Opus 5.5 (73%), Claude Sonnet 5.5 (72%), Claude Sonnet 5 (66%), averaging 70.3% on the listed scores\n- **OpenAI:** GPT-6-Sol (73%), GPT-5.6-Sol (70%), GPT-6-Luna (66%), averaging 69.7%\n- **xAI:** Grok 4.7 (69%)\n- **Alibaba:** Qwen 3.8-Max-0902 (68%)\n\nAverages come from our arithmetic on the listed scores. They cover only the models that appear in the cited results, so a vendor with a weaker model below seventh place would look better here than its full lineup deserves.\n\nAnthropic holds a slight edge on depth. Its two Claude 5.5 models sit at first and third, and its older Claude Sonnet 5 shares seventh. OpenAI matches Anthropic at the top with GPT-6-Sol and then drops to 70% and 66% for the next two entries. Readers who follow OpenAI pricing and packaging can compare these results against our report on [GPT-6.1 Terra in ChatGPT Work and Codex](https://shattered.io/gpt-6-1-terra-chatgpt-work-codex-api-pricing-2026/), since the tier a team buys affects both cost and the security profile it gets.\n\nxAI and Alibaba show that the plateau reaches beyond the two U.S. leaders. Grok 4.7 at 69% and Qwen 3.8-Max-0902 at 68% sit within five points of the top. For organizations with data-residency rules or vendor preferences, that closeness widens the field of acceptable choices. Model quality on security is converging, and no single vendor owns the category.\n\nSecurity pass rate is one axis among many. Cost, latency, context length, and agent tooling all shape which model a team uses. Our coverage of the [Gemini 4 Carbon results on a different security benchmark](https://shattered.io/gemini-4-carbon-index-59-sec-bench-pro-88-2-percent-2026/) shows how rankings shift when the test changes, which is one more reason not to lean on a single leaderboard.\n\n## Market Impact: What Engineering Leaders Will Do Next\n\nThe report changes the buying conversation in a few predictable ways. Security leaders now hold a named, vendor-specific number to put next to the productivity claims that coding assistants advertise. A 73% pass rate at the top and a 62% average give a risk officer something concrete to weigh.\n\nProcurement teams will likely ask for per-category scores. A buyer who handles payment data cares more about CWE-89 and CWE-327 than about log handling. Expect vendors to publish breakdowns by weakness class and by language, because blended averages will stop satisfying careful customers.\n\nApplication security vendors stand to benefit. When a leaderboard shows that the best generators still fail more than a quarter of the measured tasks, static analysis, dependency scanning, and runtime testing become easier to justify. Teams evaluating assistants can pair this data with our [team pilot guide for AI coding assistants](https://shattered.io/best-ai-coding-assistants-team-pilot-12-steps-2026/) to build a test that includes security from day one.\n\nCost enters the equation as well. Veracode’s Chris Wysopal noted that the top-performing models cost more. The cited material gives no prices, so we cannot quantify the premium. The trade-off is still clear: paying more for a model that fails 27% of security tasks instead of a cheaper one that fails more may or may not beat paying for scanning and review on the cheaper model’s output.\n\nInsurers and auditors may follow. Once measured security behavior of AI coding tools is public, questions about it will start appearing in vendor risk questionnaires. Companies that already track which model wrote which commit will answer those questions faster than companies that cannot.\n\n## What Chris Wysopal Said About Progress and Cost\n\nChris Wysopal, Co-founder and Chief Security Evangelist at Veracode, described the improvement as encouraging. He also pointed out two limits. The best-performing models cost more than their rivals, and they still fail more than one in four security tasks. We paraphrase his remarks here because the cited material does not provide his exact wording.\n\nBoth points deserve attention. The cost remark acknowledges that security quality is not free, and it hints at a tiering effect in which the safest code comes from the priciest models. Teams with tight budgets may settle for a cheaper model and compensate with tooling. Teams with strict compliance duties may pay for the leader and still scan everything.\n\nThe failure-rate remark matters more. A top score of 73% means the field has not reached a level at which developers can skip review. If the pass rate were 99%, an organization could debate how much oversight to keep. At 73% the debate is moot, since oversight is mandatory.\n\nNo statement from Anthropic or OpenAI appeared in the cited material. That absence is not evidence of anything. Labs often decline to comment on third-party benchmarks, and we do not read silence as agreement or disagreement.\n\n## How to Build a Secure Pipeline Around AI-Generated Code\n\nThe practical response to a 62% average is layered defense. Treat the model as a fast but fallible junior contributor whose output enters the same gates as everyone else’s. The steps below follow common application security practice and do not come from Veracode’s report.\n\n### Gate every pull request with automated analysis\n\nRun static analysis on every change, whether a human or a model wrote it. Our comparison of [SAST and DAST tools](https://shattered.io/sast-vs-dast-semgrep-sonarqube-checkmarx-snyk-codeql-zap-2026/) covers the main options and what each catches. Static scanners map well onto the four tested weakness types, since injection and unsafe cryptography leave recognizable patterns in source.\n\n### Fix the classic flaw at the source\n\nThe most frequent generated-code flaw is still string-built SQL. The snippet below shows the unsafe pattern next to the safe one. Reviewers should reject the first form on sight.\n\n```\n# Unsafe: user input is joined into the query string (CWE-89)\ncursor.execute(\"SELECT * FROM users WHERE email = '\" + email + \"'\")\n\n# Safe: the driver passes the value as a bound parameter\ncursor.execute(\"SELECT * FROM users WHERE email = %s\", (email,))\n```\n\n### Watch secrets and tooling risk\n\nModels sometimes paste credentials into example code, and developer tools themselves can become attack paths. Our [Gitleaks tutorial](https://shattered.io/gitleaks-tutorial-secret-scanning-pre-commit-github-actions-12-steps-2026/) shows how to block secrets before they reach a repository. Our report on a [tampered Claude Code extension that reached 5,000 users](https://shattered.io/cursor-tampered-claude-code-extension-5000-users-2026/) reminds teams that the assistant’s supply chain deserves scrutiny alongside the code it produces.\n\nBeyond scanning, record which model produced which change, require human review for security-sensitive modules such as authentication and payment flows, and add tests that feed hostile input to every new endpoint. None of these steps depends on which model tops the next leaderboard.\n\n## Limits of the Benchmark: What 73% Does Not Tell You\n\nAny single benchmark invites over-reading. Veracode covered four weakness categories, and real applications face dozens. A model could pass all four tested classes and still ship broken access control, insecure deserialization, or server-side request forgery. The 73% figure describes performance on a defined slice.\n\nMethodology details also shape the result. The cited material does not describe how prompts were written, whether prompts nudged models toward security, or how often tasks ran. Results from a different prompt style could rank the same models differently. A [Dark Reading piece titled So Many AI Security Benchmarks, but None Tell the Whole Story](https://www.darkreading.com/cybersecurity-analytics/benchmarking-security-of-ai-models-poses-challenges) makes the broader case that benchmark results depend heavily on design choices.\n\nReal-world use adds more variables. Developers rarely prompt a bare model for a single function. They use agents, context from existing repositories, and iterative edits, and each factor can raise or lower security outcomes. A model that scores 73% on isolated tasks might do better with a strong linter in the loop or worse inside a sprawling legacy codebase.\n\nFinally, the report comes from a company that sells application security tools. That does not make the data wrong, but readers should recognize that Veracode benefits when organizations conclude that AI code needs scanning. Outlets including TechRadar and Mezha report the leaders and the headline numbers, so the figures do not rest on a single write-up.\n\n## Five Predictions for AI-Generated Code Security\n\nThese predictions are our own analysis. Veracode has not made them, and none of them is a certainty.\n\n- **The next top score lands above 73%.** Competitive pressure and a public leaderboard give labs a reason to target security. We expect at least one model to beat the current tie in the next edition.\n- **Per-CWE reporting becomes standard.** Buyers will demand breakdowns by weakness and language, and vendors that publish them will win trust from security teams.\n- **Security pass rate joins the model card.** Labs already publish coding and reasoning scores, and a security line item is a natural addition once customers ask for it.\n- **CI gates for AI-written code spread.** Mandatory scanning on every pull request will move from best practice to baseline policy at companies that adopt coding agents widely.\n- **The average stays below 80% through the next two editions.** Averages move more slowly than peaks because older and smaller models stay in use, so a jump from 62% to 80% would need an unusual leap.\n\nEach prediction rests on the same logic. The incentive structure now rewards security gains, but the long tail of deployed models and the breadth of real-world weaknesses slow the average.\n\n## FAQ: AI-Generated Code Security\n\n**Which models topped Veracode’s AI-generated code security test?**\n\nClaude Opus 5.5 from Anthropic and GPT-6-Sol from OpenAI tied for first with a 73% security pass rate. Both recorded 100% syntax.\n\n**How many models did Veracode test, and in which languages?**\n\nVeracode tested 18 models across Python, C#, JavaScript, and Java, covering CWE-89, CWE-327, CWE-80, and CWE-117.\n\n**How much did AI code security improve?**\n\nThe average security pass rate rose from 56% to 62%, and the share of results containing vulnerabilities fell from 44% to 38%. The previous best score was 68%, set by GPT-5.5 in July 2026.\n\n**Does a 73% pass rate mean the code is safe to ship?**\n\nNo. It means the best models failed security checks on 27% of the measured tasks. Chris Wysopal of Veracode said the leading models still fail more than one in four security tasks, so scanning and human review remain necessary.\n\n**Is the Claude and GPT-6 top the leader board headline accurate?**\n\nOnly in part. The leaders are Claude Opus 5.5 and GPT-6-Sol specifically. Other models from the same families, such as Claude Sonnet 5 and GPT-6-Luna, tied for seventh at 66%.\n\n**Which model ranked third, fourth, and fifth?**\n\nClaude Sonnet 5.5 ranked third at 72% security and 99% syntax. GPT-5.6-Sol ranked fourth at 70%, and Grok 4.7 from xAI ranked fifth at 69%. Qwen 3.8-Max-0902 from Alibaba followed at 68%.\n\n**Did Anthropic or OpenAI respond to the results?**\n\nThe cited material contains no on-record comment from either company or from anyone representing them.\n\n**What should teams do about AI-generated code risk?**\n\nRun static analysis and secret scanning on every pull request, require human review for sensitive code, and track which model wrote each change. Do not treat code that compiles as code that is secure.", "url": "https://wpnews.pro/news/ai-code-security-opus-5-5-gpt-6-sol-tie-at-73-2026", "canonical_source": "https://shattered.io/ai-generated-code-security-veracode-claude-opus-5-5-gpt-6-sol-73-percent-2026/", "published_at": "2026-10-31 10:16:17+00:00", "updated_at": "2026-10-31 10:17:06.068959+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-safety", "generative-ai", "ai-research"], "entities": ["Veracode", "Claude Opus 5.5", "Anthropic", "GPT-6-Sol", "OpenAI", "GPT-5.5", "TechRadar", "Claude Sonnet 5.5"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/ai-code-security-opus-5-5-gpt-6-sol-tie-at-73-2026", "markdown": "https://wpnews.pro/news/ai-code-security-opus-5-5-gpt-6-sol-tie-at-73-2026.md", "text": "https://wpnews.pro/news/ai-code-security-opus-5-5-gpt-6-sol-tie-at-73-2026.txt", "jsonld": "https://wpnews.pro/news/ai-code-security-opus-5-5-gpt-6-sol-tie-at-73-2026.jsonld"}}