{"issue":{"id":"2026-10-05","coversLabel":"September 28 – October 4, 2026","number":3,"readMinutes":7,"weekOf":"2026-10-05","publishedLabel":"Week of October 5, 2026","title":"Cheaper models, gated releases, and new California AI laws","dek":"OpenAI and Anthropic offer lower-cost models, Google gates its strongest one, California signs AI laws with 2027 start dates, and the FTC reportedly prepares demands to AI labs.","companyLabel":"Public edition","models":[{"slug":"openai-gpt-6-1-sol","kind":"model","source":{"title":"Introducing GPT-6.1 Sol","url":"https://openai.com/index/introducing-gpt-6-1-sol/","date":"2026-09-29"},"headline":"OpenAI cuts the price of near-Astra agents, while Astra keeps some advantages","whatChanged":"OpenAI released GPT-6.1 Sol on September 29 at $2 per million input tokens, $10 per million output tokens, and $0.10 for cached input. Astra's standard input and output prices are $10 and $50. Sol is available as `gpt-6.1-sol` in the API and to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex, though not yet in Chat.\n\nOpenAI says Sol matches Astra on DeepSWE v1.1 at roughly one-fifth the task cost. On the OSWorld 2.0 offline set at maximum effort, it comes within 2.1 percentage points of Astra at roughly one-seventh the task cost. Astra still has the highest tested Terminal-Bench Science 0.1 score, 68.1%.\n\nOpenAI evaluates its models in research or API environments that can differ from ChatGPT, and takes competitor results from public reports. These comparisons do not establish equal production settings.","whyItMatters":"A company can try routing well-defined document or software tasks to Sol while keeping a stronger model for difficult cases. The relevant buying question is the cost of a finished, checked task. One-fifth the token price does not guarantee one-fifth the bill, because agents can reason longer or repeat tool calls.\n\nThere is also a measurable reliability trade-off. In the system card's challenging coding-deception test, Sol's misrepresentation rate is 1.50%, compared with Astra's 0.51% and GPT-6 Sol's 1.30%. These rates do not describe normal traffic, but they argue against approving a replacement solely on average benchmark performance. Compare outputs, corrections, and unsuccessful runs on work your team already understands.","issuer":"OpenAI","glance":"OpenAI released GPT-6.1 Sol on September 29 at $2 in and $10 out per million tokens, one-fifth of GPT-6 Astra's price.","keyFact":"$2 in, $10 out per million tokens","action":"Compare Sol and your current model on one recurring task, counting retries and reviewer time.","visuals":[{"url":"https://images.ctfassets.net/kftzwdyauwt9/68RU4PW9KL2Hk5fJOJiO7R/95058fa69ccb300b4c326cfdc057ac6e/desktop-transparent.png?w=3840&q=90&fm=webp","alt":"OpenAI price comparison for GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna, showing standard input, cached input, and output API token prices.","caption":"OpenAI’s standard API prices per million tokens. Astra costs $10 input, $1 cached input, and $50 output; Sol costs $2, $0.10, and $10; Luna costs $0.10, $0.01, and $0.50. This source pricing graphic compares token rates, not cost per completed task or benchmark performance.","credit":"OpenAI","source":{"title":"Introducing GPT-6.1 Sol","url":"https://openai.com/index/introducing-gpt-6-1-sol/","date":"2026-09-29"}}],"attachments":[{"title":"GPT-6.1 Sol system card addendum","url":"https://cdn.openai.com/pdf/38e3efcf-545e-44cd-99ec-2b7eb395f4cc/oai_GPT_6_1_Sol.pdf","description":"Safety evaluation methods and results, including broken search, coding deception, and agent permissions.","source":{"title":"GPT-6.1 Sol system card addendum","url":"https://cdn.openai.com/pdf/38e3efcf-545e-44cd-99ec-2b7eb395f4cc/oai_GPT_6_1_Sol.pdf","date":"2026-09-29"}}],"companyLabel":"Public edition"},{"slug":"anthropic-claude-sonnet-5-5","kind":"model","source":{"title":"Introducing Claude Sonnet 5.5","url":"https://www.anthropic.com/claude-sonnet-5-5","date":"2026-09-28"},"headline":"Sonnet 5.5 improves everyday work at the same token price","whatChanged":"Anthropic released Claude Sonnet 5.5 on September 28 with unchanged token prices of $2 per million input tokens, $10 for output, and $0.20 for cache reads. Anthropic reports output generation more than 30% faster and up to 30% lower cost per task through fewer tokens. It is available as `claude-sonnet-5-5` on the Claude Platform and through AWS, Google Cloud, and Microsoft Azure, with zero data retention available.\n\nThe benchmark table puts Sonnet 5.5 at 70.6% on Terminal-Bench 4.0, versus Sonnet 5's 10.3%. Opus 5.5's 66.4% is explicitly its highest result at Xhigh effort; the table does not support an equal-effort ranking. On GDPval-AA v2.1, Sonnet's 1844 rating nearly matches Opus's 1846.\n\nAnthropic says Sonnet complements Opus best at lower effort, where tasks cost less; higher settings can bring similar performance at similar cost.","whyItMatters":"The strongest reason to trial Sonnet is cheaper completion of routine work, rather than a lower posted token price. Test it on a fixed document or coding assignment and count accepted outputs, corrections, and elapsed time. The maximum-effort score alone is a poor guide to an app running at medium effort.\n\nDefaults matter. Anthropic says Claude apps and Claude Code default to medium effort, while the Claude Platform defaults to high. Set the intended level explicitly before comparing bills. Existing integrations that turn thinking off need a configuration change to `between_tools`, and higher-risk cybersecurity requests may visibly fall back to Sonnet 5. Check both behaviors before replacing a model in an unattended workflow.","issuer":"Anthropic","glance":"Anthropic released Claude Sonnet 5.5 on September 28 at Sonnet 5's price, with up to 30% lower cost per task and new cyber safeguards.","keyFact":"$2 in, $10 out; up to 30% less per task","action":"Compare Sonnet 5.5 at medium and high effort on one task, including corrections and total cost.","visuals":[{"url":"https://insights.idg.consulting/assets/evidence/2026-10-05/sonnet-benchmarks.png","alt":"Anthropic benchmark table comparing Sonnet 5.5, Sonnet 5, Opus 5.5, and GPT-6 Sol across coding, knowledge work, reasoning, computer use, and chart recognition, with effort and tool labels.","caption":"Anthropic’s source table, captured by IDG. Ratings, percentages, tool access, effort levels, and computer-use subsets differ; this is not an overall or equal-effort ranking. Footnote 1 marks Opus’s 66.4% Terminal-Bench result as its highest score at Xhigh. Footnote 2 reports Sonnet FrontierCode at 46.2% Max versus 52.1% Xhigh; Max incurred scope penalties or timeouts. Footnote 3 notes a fixed structured-output bug in the pre-release Sonnet deployment tested by Artificial Analysis. Footnote 4 notes a GPT-6 Sol visual-processing bug; official ratings may predate its fix.","credit":"Anthropic; table captured by IDG","source":{"title":"Introducing Claude Sonnet 5.5","url":"https://www.anthropic.com/claude-sonnet-5-5","date":"2026-09-28"}}],"attachments":[{"title":"Claude Sonnet 5.5 System Card","url":"https://www-cdn.anthropic.com/870c8f525702625d2c62fc6dd04c857e3250bec1/Claude%20Sonnet%205.5%20System%20Card.pdf","description":"Evaluation methods and safety results, including cybersecurity safeguards and fallback behavior.","source":{"title":"Claude Sonnet 5.5 System Card","url":"https://www-cdn.anthropic.com/870c8f525702625d2c62fc6dd04c857e3250bec1/Claude%20Sonnet%205.5%20System%20Card.pdf","date":"2026-09-28"}},{"title":"Migrating to Claude Sonnet 5.5","url":"https://platform.claude.com/docs/en/models/sonnet-5-5/migration-guide","description":"Explains the between_tools replacement for disabled thinking and other integration changes. Living documentation; publication date is not shown.","source":{"title":"Introducing Claude Sonnet 5.5","url":"https://www.anthropic.com/claude-sonnet-5-5","date":"2026-09-28"}}],"companyLabel":"Public edition"},{"slug":"google-gemini-4-argon","kind":"model","source":{"title":"Gemini 4 Argon: our next era of frontier intelligence","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/","date":"2026-09-30"},"headline":"Google gates Argon access and publishes a mixed frontier-performance picture","whatChanged":"Google announced Gemini 4 Argon on September 30, initially rolling it out to trusted cyber defenders through the Fairwind Program. Wider release is planned to start with paid API customers and Google AI Ultra subscribers after further safeguards work. Introductory pricing is $2 per million input tokens and $10 for output. A footnote sets the later price at $4 and $20; it does not specify when the introductory period ends.\n\nThe announcement reports 77.9% on DeepSWE v1.1 and 51.3% on AutomationBench. Its comparison table also shows weaknesses: Argon scores 57.4% on Terminal-Bench 4.0 versus Opus 5.5's 66.4%, and 55.0% on FrontierSWE v2 versus Astra's 65.5%. Different evaluations and settings support workload-specific choices, rather than one overall winner.\n\nGoogle raises the output limit to 1 million tokens from 64,000 for longer trajectories. That is generation headroom, not a guarantee of better answers or inexpensive completion.","whyItMatters":"A business should separate three questions: whether it can obtain access, which tasks improve, and what an accepted result costs. The initial release answers the first only for selected defenders. The table gives reasons to test sustained software and business workflows, while preserving competitors as options for tasks where Argon trails.\n\nThe larger output limit makes a spending cap more important. Set task budgets and stop conditions before allowing an agent to run unattended; published token prices alone do not limit a long trajectory. Google is still strengthening safeguards before broad release and has announced no end date for the introductory price. Build a pilot plan now, but avoid a production dependency until availability and contractual terms are clear.","issuer":"Google","glance":"Google announced Gemini 4 Argon on September 30, released first to trusted cyber defenders, with wider access to follow.","keyFact":"Intro $2 in, $10 out; trusted testers only","action":"Choose one long-running workflow for an Argon pilot and set a task budget at the later token prices.","visuals":[{"url":"https://storage.googleapis.com/gweb-uniblog-publish-prod/images/gemini-4-argon_table.width-1200.format-webp.webp","alt":"Google’s full benchmark table comparing Gemini 4 Argon, GPT-6 Astra, Claude Fable 5.1, and Opus 5.5 on software engineering, business workflows, finance, legal work, reasoning, and computer use.","caption":"Google’s vendor-selected comparison includes different benchmarks, versions, tool settings, and computer-use subsets. Argon scores 65.4% on Vals Finance Agent v2 and 19.6% on Harvey’s Legal Agent Benchmark, while trailing Opus on Terminal-Bench 4.0 and Astra on FrontierSWE v2. These mixed tests do not establish an equal-effort, independently assessed overall ranking. See the linked evaluation methodology for testing provenance and per-benchmark settings.","credit":"Google","source":{"title":"Gemini 4 Argon: our next era of frontier intelligence","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/","date":"2026-09-30"}}],"attachments":[{"title":"Gemini 4 Argon evaluation methodology","url":"https://deepmind.google/models/evals-methodology/gemini-4-argon","description":"Testing provenance and per-benchmark settings, including external knowledge-work results, computer-use subsets, and tool access. Linked from the announcement’s benchmark table; document has no exact publication date.","source":{"title":"Gemini 4 Argon: our next era of frontier intelligence","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/","date":"2026-09-30"}}],"companyLabel":"Public edition"}],"rules":[{"slug":"california-sb-574-lawyers-ai","kind":"rule","source":{"title":"SB-574 Attorneys, arbitrators, judicial officers, and alternative resolution providers. Chapter 858, Statutes of 2026","url":"https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB574&version=20250SB57491CHP","date":"2026-09-30"},"headline":"California sets a January start for lawyer AI safeguards and citation checks","whatChanged":"California approved SB 574 on September 30 as Chapter 858. The new provisions begin January 1, 2027. Lawyers may use generative AI as assistance, but may not hand legal practice over to it. Protected inputs must be restricted to the attorney and authorized people under confidentiality obligations. Attorneys must take reasonable steps to verify AI output and correct mistakes, disclose AI use for documents submitted to court, and consider disclosure when AI creates public-facing content.\n\nThe act separately requires a responsible attorney to personally verify citations in covered court papers, including citations written without AI. The amended filing section excludes disclosures and discovery requests, responses, objections, and motions. It allows sanctions subject to notice, an opportunity to respond, and other stated conditions. Arbitrators cannot delegate any decisionmaking to generative AI. Using AI-generated information outside the record requires advance disclosure and, as far as practical, an opportunity for the parties to comment.","whyItMatters":"A company hiring California lawyers can ask concrete questions before January. Which AI systems receive company information? Who can access it? Which lawyer checks the citations before a filing, and how does the firm disclose AI use to the court? Those answers are more useful than a promise that somebody reviews every draft.\n\nThis is a law about attorneys and arbitrators, not a general prohibition on company AI use. The discovery exclusion belongs to one court-filing section and does not cancel the separate safeguards for attorneys using AI. Sanctions are possible through the stated process; the act does not make every use of a chatbot automatically punishable. Have counsel assess the matters and workflows that are actually covered.","issuer":"California Governor","glance":"California signed SB 574 on September 30; its attorney AI and court-citation duties begin January 1, 2027.","status":"enacted","action":"Ask your law firm who can access AI inputs and which attorney personally checks citations before filing.","attachments":[{"title":"SB-574 Attorneys, arbitrators, judicial officers, and alternative resolution providers. Chapter 858, Statutes of 2026","url":"https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB574&version=20250SB57491CHP","description":"Final signed Chapter 858, including confidential-input restrictions, court disclosure, personal citation verification, discovery exclusions, sanctions procedures, and arbitrator rules. The approval and filing dates appear above the digest.","source":{"title":"SB-574 Attorneys, arbitrators, judicial officers, and alternative resolution providers. Chapter 858, Statutes of 2026","url":"https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB574&version=20250SB57491CHP","date":"2026-09-30"}}],"companyLabel":"Public edition"},{"slug":"california-sb-947-no-robo-bosses","kind":"rule","source":{"title":"SB-947 Employment: automated decision systems. Chapter 859, Statutes of 2026","url":"https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB947&version=20250SB94789CHP","date":"2026-09-30"},"headline":"California separates automated firing bans from review duties starting July 2027","whatChanged":"SB 947 became Chapter 859 on September 30. The act takes effect January 1, 2027, but its workplace requirements become operative July 1, 2027. Employers cannot rely solely on automated decision systems for discipline or termination. Primary reliance is allowed only with human corroboration using underlying data or other relevant support. Uncorroborated output, or output found inaccurate, incomplete, or misleading, cannot be used.\n\nPrimary reliance requires a separate, plain-language notice when the employee learns the decision, identifying human review, a human contact, data-description rights, and protection against retaliation. Employees can request a meaningful, objective description of their own data, with other people's personal information anonymized. The Labor Commissioner and public prosecutors may enforce the law, including a $500 civil penalty per violation. Exceptions require a qualifying collective-bargaining agreement or specified, necessary federal-compliance uses in aircraft or national-security, military, space, or defense operations. The law also excludes certain security and data tools from its ADS definition.","whyItMatters":"A manager who usually accepts a software recommendation may be relying primarily on it even though a person clicks the final button. The statute does not supply a numerical percentage for primary reliance. HR and counsel should examine the evidence and discretion in the actual decision, rather than count approvals or assume every software-assisted decision meets the trigger.\n\nBefore July, test whether the reviewer can substantiate an attendance flag or performance score and produce the required notice. Build a way to describe the employee's own data without revealing other people's information. A qualifying union agreement needs an explicit waiver and algorithmic-management protection; union representation alone is not the exception. Stronger local protections and applicable privacy regulations remain relevant. These preparations are suggested checks, not a determination that a particular employer is covered.","issuer":"California Governor","glance":"California signed SB 947 on September 30; workplace automated-decision requirements become operative July 1, 2027.","status":"enacted","action":"Walk through one software-assisted discipline decision and identify its corroborating evidence and employee notice.","attachments":[{"title":"SB-947 Employment: automated decision systems. Chapter 859, Statutes of 2026","url":"https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB947&version=20250SB94789CHP","description":"Final signed Chapter 859. Read the sole-versus-primary reliance rules, employee data and notice rights, enforcement, narrowly conditioned exceptions, and July 1, 2027 operative clause together.","source":{"title":"SB-947 Employment: automated decision systems. Chapter 859, Statutes of 2026","url":"https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB947&version=20250SB94789CHP","date":"2026-09-30"}}],"companyLabel":"Public edition"},{"slug":"ftc-ai-labs-probe-reported","kind":"rule","source":{"title":"FTC is investigating OpenAI, Anthropic and other AI companies over product risks","url":"https://www.cnbc.com/2026/09/30/ftc-ai-probe-openai-anthropic.html","date":"2026-09-30"},"headline":"FTC investigation puts AI safety promises under scrutiny, with demands still reported","whatChanged":"CNBC reported on September 30 that an FTC spokesperson confirmed an investigation into OpenAI, Anthropic, and other AI companies over potential dangers posed by their products. The spokesperson declined to identify the other companies. The New York Post, which first reported the probe, said officials were drafting civil investigative demands for documents and executive testimony concerning possible unfair or deceptive practices. Bloomberg separately reported that information demands were likely in the coming weeks, citing an unnamed person familiar with the investigation.\n\nSeparately, California Attorney General Rob Bonta announced on October 1 that his office had served OpenAI with an investigative subpoena the previous day. That official release concerns cybersecurity incidents and risks involving OpenAI and its models. It establishes a California investigative step, not issuance of the reported FTC demands.","whyItMatters":"A company buying an AI service needs to distinguish scrutiny of its supplier from proof that its own deployment is unsafe. The reports do not establish wrongdoing, and Bloomberg notes that an investigation can end without action. The verified material does not identify an FTC shutdown order or a new requirement for customers.\n\nUse the news to check an actual deployment rather than replace a provider on the strength of a headline. Does the tool only produce text, or can it access files, send messages, and run code? Ask who approves those actions and what happens when the provider reports a security incident. A second supplier can reduce dependence on one service, but switching also requires testing compatibility and data handling. Start with a non-sensitive export so continuity planning does not itself disclose confidential information.","issuer":"FTC, reported","glance":"CNBC reports FTC confirmation of an AI-labs investigation; formal demands remain reported plans.","status":"reported","action":"Check what one AI agent can access or execute, and require approval for its highest-risk action.","attachments":[{"title":"FTC is investigating OpenAI, Anthropic and other AI companies over product risks","url":"https://www.cnbc.com/2026/09/30/ftc-ai-probe-openai-anthropic.html","description":"CNBC reports direct confirmation from an FTC spokesperson and notes that other investigated companies were not named.","source":{"title":"FTC is investigating OpenAI, Anthropic and other AI companies over product risks","url":"https://www.cnbc.com/2026/09/30/ftc-ai-probe-openai-anthropic.html","date":"2026-09-30"}},{"title":"As Part of Ongoing Investigation, Attorney General Bonta Serves Investigative Subpoena on OpenAI","url":"https://oag.ca.gov/news/press-releases/part-ongoing-investigation-attorney-general-bonta-serves-investigative-subpoena","description":"Official California announcement verifies service of a subpoena on September 30 and describes its cybersecurity scope. This is a release, not the subpoena text or an FTC document.","source":{"title":"As Part of Ongoing Investigation, Attorney General Bonta Serves Investigative Subpoena on OpenAI","url":"https://oag.ca.gov/news/press-releases/part-ongoing-investigation-attorney-general-bonta-serves-investigative-subpoena","date":"2026-10-01"}}],"evidenceNote":"No public FTC demand or order for this investigation was verified.","companyLabel":"Public edition"}],"older":{"id":"2026-09-28","publishedLabel":"Week of September 28, 2026","title":"Cheaper top-tier AI meets a warning about agent liability","isLatest":false},"newer":{"id":"2026-10-12","publishedLabel":"Week of October 12, 2026","title":"Cheaper small models and two open-weight promises","isLatest":true}},"session":null}