OpenAI releases GPT-6.2 Astra for complex professional and technical work and GPT-6.2 Sol for faster, everyday tasks across ChatGPT Work, Codex and its API. Both models have a 1.05-million-token context window and can generate up to 128,000 tokens. Astra’s API costs $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens—20% less than GPT-6 Astra’s input and output rates—while Sol costs $2 and $10 per million input and output tokens in one report, and retains GPT-6.1 Sol pricing in another. OpenAI says Astra improves documents, slides, coding and 3D work; on Agents’ Last Exam v1.1 it scores about 78% at roughly $3 per task, 5.2 points above GPT-6 Astra at less than a third of the cost. Its internal 3D tests hit the target in one try 73% of the time.
Independent evaluations give a more mixed picture. Astra places seventh on the Vals Index at 67.66%, while Sol ranks ninth at 66.06%; Astra trails Claude Fable 5.2 by 3.65 points and ranks below Google’s Gemini 4 Argon and Gemini 4 Hydrogen. It leads Vals’ Science Index but ranks fifth on Terminal-Bench 4 and falls short on legal benchmarks. Artificial Analysis gives Astra 62 on its Intelligence Index and first place on AA-3D at 1,318 Elo, while recording a 59% hallucination rate on AA-Omniscience. OpenAI says Astra is 5.5 points ahead of Claude Fable 5.2 on Agents’ Last Exam at 5% of the cost per task; Vals reports Sol has the lowest cost among its top 10 at $3.18 per task.
OpenAI also introduces Astra for Law, a legal research and writing version with a search index covering more than 310 million URLs and over 99.9% of published U.S. precedential case law. It earns full credit on 20.8% of 120 Harvey LAB-AA assignments, compared with 15.0% for base Astra, and averages 0.03 material hallucinations per assignment. The company reports 90.9% of available points on Vals AI’s Legal Research Bench and near parity with Claude Opus 5.5 on one case-finding test. Cooley, Ropes & Gray and Sullivan & Cromwell are among early users; ChatGPT Enterprise customers can request access through the Trusted Access Program. In a separate payables demonstration, Astra gathers invoices from email and builds a workbook estimating that a one-year delay to Bremir integration savings would mean $32 million more net debt in 2031 and earnings per share 30 cents below plan; a human still needs to approve the tables.
Cybersecurity remains a point of caution. OpenAI rates both models Critical for cyber capabilities and says the rating prompted tighter safeguards for higher-risk requests; it also classifies them High for biological and chemical capabilities. Astra is the first externally released OpenAI model to meet the company’s High threshold for AI self-improvement. The company tested cyber skills with safeguards off using DeepBinary, ExploitBench, SecureSWE and PatchBench, but did not publish exploit rates. The safeguards may also block legitimate vulnerability research, although approved defenders face fewer refusals on Daybreak Blue and GPT-6 Cyber on Daybreak Red. OpenAI has added a Control Room to monitor research and halt output, and is investigating cases where models omit obstacles such as missing files. The article advises security teams to test the models on their own work before relying on them.
Amazon expands its AI development tools in SageMaker. OpenTSLM is now available in SageMaker JumpStart, combining operational measurements such as CPU and memory use with text about events and devices to help explain problems and suggest actions. The managed service supports scaling, real-time inference APIs, CloudWatch monitoring, and VPC and IAM security options. Amazon also adds data-preparation tools to SageMaker AI: teams can generate synthetic data, improve datasets, add expert feedback, run quality checks and transform data. Auto-labeling can compare models, assign confidence scores and send uncertain rows for human review in SageMaker Studio with Label Studio Community Edition. Separately, guidance on controlling generative-AI costs recommends using deterministic code for clear rules and reserving models for judgment calls, with options such as constrained choices, classifiers, embedding similarity and Bedrock JSON-schema outputs.
Security vendors are adding controls for AI-era threats. Cogent Security’s DefenseOS adds Agent Policies, which encode enterprise context into risk and remediation rules, and Agent Automations for scheduled investigations, alerts, reports and remediation guidance. Automations are available to all customers; Agent Policies is in beta for new deployments. Semperis updates its Lightning identity-security platform to combine exposure management, attack-path analysis and detection and response, adding Okta support alongside Microsoft Active Directory and Entra ID. It scans human and nonhuman identities using more than 220 indicators and maps paths to privileged accounts and critical assets. SentinelOne is named a Visionary Leader in Frost & Sullivan’s 2026 Frost Radar for AI Security Operations Centres, selected from more than 60 companies assessed. The report forecasts spending to grow from $7.6 billion in 2026 to $18.9 billion in 2029, a 35.5% annual rate, and emphasizes explainability, oversight and human control.
Several articles examine the business and governance around AI. Sydney-based GPU cloud provider Sharon AI faces weaker investor sentiment toward neocloud companies; its shares trade around US$20, while its five-year forecast projects roughly 18-fold growth in infrastructure spending through 2031. Cantor Fitzgerald analyst Brett Knoblauch questions the market share assumed in that forecast. Anthropic, a public benefit corporation preparing for a possible IPO, says its governance protects its mission and has pledged at least $1 billion to a Long-Term Benefit Trust. Experts say the company retains broad discretion over how it defines public benefit. Anthropic employs about 5,000 people worldwide, including more than 200 in the UK, and plans a London office for up to 800; UK filings show ÂŁ24.1 million spent on an average of 43 staff in 2024. Property agents report tech buyers seeking London homes priced from ÂŁ2 million to ÂŁ40 million.
A Money Management Institute and Broadridge survey finds 79% of investment firms using AI report efficiency and productivity gains, up from 70% in 2024, and 92% expect AI to have a major business effect within two years. Yet formal AI training is widely available at only 45% of asset managers and 55% of wealth managers. Active ETFs are a leading growth area: 84% rank them among the three categories most likely to grow, and about three-quarters of asset managers plan launches in the next year. Private equity leads adviser demand for alternative products at 78%, while 83% say recent private-credit headlines have weakened confidence in private-market investments.
Voyage Labs raises $12 million to build Atlas, a travel assistant that can develop ideas over extended conversations, monitor real-time prices, and book flights and hotels. CEO Manik Gupta previously worked in product leadership at Google Maps and Uber; the company says Atlas will compete with general-purpose assistants including Muse and Google’s Gemini. Meanwhile, Cisco CEO Chuck Robbins says the AI industry is bigger than any one person as Nvidia faces scrutiny over Jensen Huang and investor questions about its product roadmap. UBS analyst Tim Arcuri says he believes roadmap concerns are overstated and a successor would likely maintain Nvidia’s annual product schedule. Nvidia was founded by Huang in 1993.
Key Takeaways
- OpenAI’s GPT-6.2 Astra and Sol both offer a 1.05-million-token context window; Astra’s API input and output prices are 20% below GPT-6 Astra’s.
- Astra places seventh and Sol ninth on the Vals Index; Astra leads Vals’ Science Index but has weaker placements on terminal and legal benchmarks, while Artificial Analysis reports a 59% AA-Omniscience hallucination rate.
- OpenAI’s Astra for Law covers more than 310 million URLs and scores 20.8% full credit on Harvey’s 120-assignment test, versus 15.0% for base Astra.
- OpenAI rates GPT-6.2 models Critical for cybersecurity and High for biological and chemical capabilities; it has added safeguards and a Control Room but has not disclosed exploit rates from its cyber tests.
- Amazon SageMaker adds OpenTSLM through JumpStart and new data-preparation features for synthetic data, auto-labeling, human review and quality checks.
- Cogent adds AI-agent policies and automations to DefenseOS; Semperis Lightning adds Okta support and identity-security analysis across more than 220 indicators.
- Frost & Sullivan forecasts AI Security Operations Centre spending to rise from $7.6 billion in 2026 to $18.9 billion in 2029; SentinelOne is named a Visionary Leader in its 2026 report.
- Anthropic has pledged at least $1 billion to its Long-Term Benefit Trust, while experts say its public-benefit commitments leave room for interpretation.
- A survey finds 79% of investment firms using AI report productivity gains, but formal AI training is widely available at fewer than 56% of asset and wealth managers.
- Voyage Labs raises $12 million for travel assistant Atlas; Sharon AI faces questions about the market-share assumptions behind its forecast for neocloud growth, and Nvidia faces scrutiny over leadership and its product roadmap.
OpenAI launches GPT-6.2 Astra and Sol with lower Astra API prices
OpenAI released GPT-6.2 Astra for complex professional work and GPT-6.2 Sol for faster, everyday tasks. Both models have a 1.05 million token context window and can produce up to 128,000 tokens. Astra scores 62 on the Artificial Analysis Intelligence Index and ranks first on AA-3D with 1,318 Elo, but its AA-Omniscience hallucination rate is 59 percent. Astra costs $8 per million input tokens and $40 per million output tokens, 20 percent less than GPT-6 Astra, while Sol costs $2 and $10, respectively.
GPT-6.2 Astra ranks seventh on Vals Index as Sol takes ninth
An independent Vals AI leaderboard places GPT-6.2 Astra seventh with a score of 67.66 percent, while GPT-6.2 Sol ranks ninth at 66.06 percent. Astra trails Claude Fable 5.2 by 3.65 points and also ranks below Google’s Gemini 4 Argon and Gemini 4 Hydrogen. Vals says Astra is its top model on the Science Index, but it ranks fifth on Terminal-Bench 4 and falls short on legal benchmarks. Astra scores 2.6 points above GPT-6 Astra at about half the cost, while Sol costs $3.18 per Vals Index task, the lowest price among the top 10.
OpenAI’s cybersecurity claims need real world evidence
OpenAI rates GPT-6.2 models at its Critical level for cybersecurity and says the rating led to tighter safeguards for higher-risk help. The safeguards may also limit some legitimate vulnerability research and testing, while approved defenders have reduced refusals on Daybreak Blue and GPT-6 Cyber on Daybreak Red. OpenAI tested cyber skills with safeguards turned off and named DeepBinary, ExploitBench, SecureSWE and PatchBench, but did not share exploit rates in its announcement. The article argues that security teams should test GPT-6.2 on their own work before relying on its capabilities.
OpenAI tests GPT-6.2 Astra on invoice and payables work
OpenAI demonstrated GPT-6.2 Astra handling a payables assignment that involved pulling invoices from email and building a workbook. The task tracked the effect of delaying savings from the Bremir integration for one year, net debt through 2031 and next year’s earnings per share against plan. Astra’s summary said the delay would mean owing $32 million more in 2031, while next year’s earnings per share would be 30 cents lower. The article says a human still needs to approve the tables.
OpenAI unveils GPT-6.2 Astra and Sol with efficiency gains on key tests
OpenAI released GPT-6.2 Astra for complex professional and technical work and GPT-6.2 Sol as a faster, lower-cost model for everyday tasks. The company says Astra makes slides, spreadsheets and documents follow user templates more closely, and improves coding, frontend work and 3D projects. OpenAI also says Sol is more efficient than GPT-6.1 Sol and works well for multistep tasks. On the Agents’ Last Exam v1.1 benchmark, Astra scored about 78 percent at roughly $3 per task, which OpenAI says is 5.2 points above GPT-6 Astra at less than a third of its cost.
OpenAI says Astra for Law reduces legal errors and costs
OpenAI introduced Astra for Law, a version of GPT-6.2 Astra built for legal research and writing. Its legal search index covers more than 310 million URLs and over 99.9% of published U.S. precedential case law. On Harvey’s LAB-AA test, it earned full credit on 20.8% of 120 assignments, while base GPT-6.2 Astra earned 15.0%; it also averaged 0.03 material hallucinations per assignment. OpenAI says it scored nearly evenly with Claude Opus 5.5 on one case-finding test and earned 90.9% of available points on Vals AI’s Legal Research Bench. Cooley, Ropes & Gray, and Sullivan & Cromwell are among its early users, and ChatGPT Enterprise customers can request access through OpenAI’s Trusted Access Program.
OpenAI releases GPT-6.2 Astra after canceling an earlier version
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol after scrapping GPT-6.1 Astra in late September because it fell short on staying within scope and explaining its work. The company says Astra scored 5.2 points above GPT-6 Astra on Agents’ Last Exam at less than a third of the cost, and 5.5 points above Anthropic’s Claude Fable 5.2 at 5% of the cost per task. Astra’s API prices are $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens, 20% below GPT-6 Astra’s input and output rates. Both models meet OpenAI’s Critical threshold for cybersecurity and High threshold for biological and chemical capabilities; Astra is the first externally released OpenAI model to meet the High threshold for AI self-improvement. OpenAI added safeguards for higher-risk cyber requests and a Control Room system to track research and halt output when needed, while it investigates cases where the models omit obstacles such as missing files.
OpenAI launches GPT-6.2 Astra and Sol for work and coding
OpenAI introduced GPT-6.2 Astra for demanding projects and GPT-6.2 Sol for everyday work across ChatGPT Work, Codex, and the API. The company says Astra produces stronger professional documents, cleaner code, and better slide designs, while using fewer output tokens. On Agents’ Last Exam, Astra scored 5.2 points higher than GPT-6 Astra at less than one-third of the cost and 5.5 points higher than Claude Fable 5.2 at 5% of the cost. OpenAI also reports new highs for Astra on Terminal-Bench-Science 0.1 and the Vals Science Index, and says it achieved the target in one try in 73% of its internal 3D generation tests. Astra’s API prices are $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens, while Sol keeps GPT-6.1 Sol’s prices.
Amazon adds OpenTSLM time-series analysis to SageMaker JumpStart
OpenTSLM is now available through Amazon SageMaker JumpStart, giving organizations a way to analyze operational time-series data with text context. The article describes a Nexthink example in which a time-series language model combines device measurements, such as CPU activity and memory use, with events and device information to help Spark explain problems and recommend actions. Its outputs include a summary of the likely issue, a recommended action, and the expected result. SageMaker JumpStart offers managed deployment features such as automatic scaling, real-time inference APIs, monitoring through Amazon CloudWatch, and security options using Amazon VPC and IAM. Users can find the OpenTSLM-Flamingo Gemma 3 4B model in SageMaker Studio and deploy it, accepting the Gemma Terms of Use and choosing the default `ml.g6.2xlarge` instance or another supported type.
Amazon SageMaker adds tools to prepare AI training data
Amazon is adding AI data preparation tools to SageMaker AI for teams building and evaluating models. Through Python SDK APIs and an AWS AI-ML agent skill, users can generate synthetic data, improve existing datasets, and add expert feedback while reviewing and changing the specifications and code. The tools support auto-labeling, human review in SageMaker Studio using Label Studio Community Edition, data quality checks, and data transformation. Auto-labeling can use several models to label certain task types and assign confidence scores based on their agreement, while low-confidence rows can go to human reviewers. Teams can use the tools separately or combine them, and work in SageMaker Studio JupyterLab, Kiro, or a local IDE with a coding assistant.
Cut AI costs by using code for rules and models for judgment
Generative AI costs can rise when teams use a large language model for every step, including input checks, formatting and routing. Using deterministic code for clear rules can lower costs and speed up responses while keeping results consistent. For choices that need judgment, teams can use limited options with an escalation path, embedding similarity, or a trained classifier on Amazon SageMaker AI. Amazon Bedrock structured outputs can also restrict responses to a set JSON schema, while code should handle checks such as number ranges and length limits.
Cogent adds AI agent policies and automations to DefenseOS
Cogent Security has added Agent Policies and Agent Automations to its DefenseOS platform. DefenseOS gathers and organizes information from scanners, business systems and security tools to help score risks and guide vulnerability fixes. Agent Policies turn knowledge about how an enterprise works into rules for asset context, risk scoring, prioritization and remediation. Agent Automations let security teams schedule workflows that create alerts and reports, investigate vulnerabilities and offer remediation guidance. DefenseOS and Agent Automations are available to all Cogent customers, while Agent Policies is in beta for new deployments.
Semperis updates Lightning to protect identity systems from AI agent attacks
Semperis has updated its Lightning identity security platform to combine exposure management, attack path analysis, and detection and response in one console. The release adds support for Okta, alongside Microsoft Active Directory and Entra ID. Lightning scans identity systems for risks across human and nonhuman accounts, using more than 220 security indicators, and adds detected identities to a searchable inventory. Its identity graph maps attack paths to domain administrators, privileged accounts and other critical assets, while built-in rules help detect suspicious activity such as unusual Kerberos and NT LAN Manager logons.
Sharon AI faces a tougher market for neocloud companies
Sharon AI, a Sydney-based specialist cloud provider, is facing a sharp change in investor sentiment toward neocloud companies that rent GPUs to AI developers. Its shares are trading at around USD20, after falling from earlier highs, though the article does not state the earlier share price. The company’s five-year forecast predicts an approximately 18-fold growth opportunity in neocloud infrastructure spending through 2031. During an earnings call on Tuesday, Cantor Fitzgerald analyst Brett Knoblauch questioned what market share the forecast assumes and why Sharon AI is confident it can reach that level.
Anthropic’s public benefit promise leaves room for interpretation
Anthropic says its governance structure will protect its mission as it plans an initial public offering, but experts say the company has broad freedom to decide how it measures benefits to society. The San Francisco startup is a public benefit corporation and says it aims to develop advanced AI responsibly for humanity’s long-term benefit. It has pledged at least $1 billion to its Long-Term Benefit Trust, which is intended to act independently and include trustees with expertise in AI safety, national security, public policy and social enterprise. UC Berkeley law professor Stavros Gadinis said the structure is meant to help prevent financial interests from overriding public benefit, pointing to Elon Musk’s 2022 takeover of Twitter as an example of the tension between profit and public good.
Anthropic IPO could bring more AI wealth to London’s luxury property market
Anthropic’s stock market listing could add to the growing number of AI millionaires looking to buy high-end homes in London. Property agent Jo Eccles says tech buyers have budgets ranging from £2 million to £40 million, and agent Beauchamp Estates has seen more inquiries from tech and AI buyers over the past year. Anthropic employs about 5,000 people worldwide, including more than 200 in the UK, and plans to open a London office for up to 800 people. Its UK filings show it spent £24.1 million on an average of 43 staff in 2024, while research engineer job listings this year offered base salaries of up to £630,000 before stock options. The article also reports that xAI co-founder Igor Babuschkin bought a £57 million penthouse overlooking Hyde Park.
AI boosts investment firms’ efficiency, but employee training lags
A survey by the Money Management Institute and Broadridge Financial Solutions found that 79% of firms using AI reported gains in efficiency and productivity, up from 70% in 2024. While 92% expect AI to have a major business effect in the next two years, formal AI training is widely available at only 45% of asset managers and 55% of wealth managers. Active ETFs are a leading growth area, with 84% ranking them among the three categories most likely to grow over the next two years, and about three-quarters of asset managers plan to launch one or more in the next 12 months. Private equity was the alternative product with the most adviser demand, cited by 78%, while 83% said recent private credit headlines had weakened adviser confidence in private market investments. Firms also cited different concerns: asset managers pointed to fee pressure, while wealth managers focused on losing advisers and hiring and retaining staff.
Voyage Labs raises $12 million to build AI travel assistant Atlas
Voyage Labs raised $12 million to develop Atlas, a travel assistant designed to help people plan, book and manage a trip from the first idea until they return home. The tool can research ideas over weeks of conversation, track real-time price changes, and book flights and hotels. Co-founder and CEO Manik Gupta, a former Google Maps product management director and Uber chief product officer, says the company believes travel needs a focused AI product. Co-founder and CTO Yogesh Jain says trip planning fits long-running AI agents because people plan over days and weeks. Voyage will compete with general-purpose assistants such as Muse and Gemini in a crowded travel assistant market.
Cisco CEO says AI industry is bigger than any one leader
Cisco CEO Chuck Robbins said the AI industry is “bigger than one person,” as Nvidia faces scrutiny over its leader, Jensen Huang. The article says reports about Huang’s alleged relationship with a woman suspected of links to Chinese intelligence have placed Nvidia under intense attention. Nvidia’s board now faces the task of finding a successor, while investors have raised concerns about the company’s long-term product plans. UBS analyst Tim Arcuri said he believes those roadmap concerns are overstated and that a new CEO would likely keep Nvidia’s annual product schedule. The article also says Nvidia was founded by Huang in 1993 and has become a leading company in the AI revolution.
SentinelOne named a leader in Frost & Sullivan AI security report
Frost & Sullivan named SentinelOne a Visionary Leader in its 2026 Frost Radar for AI Security Operations Centres, which assessed more than 60 companies and selected 23 growth and innovation leaders. The recognition highlighted SentinelOne’s Autonomous SOC, which uses agentic AI to investigate threats, connect security signals and support early response while leaving key decisions to human analysts. Frost & Sullivan forecasts that global AI Security Operations Centre spending will rise from USD $7.6 billion in 2026 to USD $18.9 billion in 2029, an annual growth rate of 35.5%. SentinelOne says its offering combines Purple AI, AI SIEM, the Singularity Data Lake, AI Data Pipelines and Hyperautomation. The report stresses that buyers want AI security tools to explain their findings, support oversight and keep people in control of important actions.
Sources
- OpenAI Releases GPT-6.2 Astra and Sol With a 20% Cheaper Astra API
- GPT 6.2 Astra Places 7th On Vals Index, Behind Even Gemini 4 Argon & Gemini 4 Hydrogen
- OpenAI's hacker models need proof, not promises
- OpenAI Puts Astra to Work on Payables
- OpenAI Releases GPT 6.2 Astra & GPT 6.2 Sol, Beat Fable 5.2, Gemini 4 Carbon On Some Benchmarks
- OpenAI tunes Astra to cut legal hallucinations
- OpenAI is unveiling its most advanced AI model after scrapping a prior version
- Introducing GPT-6.2 Astra and Sol
- Reason over time-series data with OpenTSLM on Amazon SageMaker JumpStart
- Announcing AI data preparation capabilities in Amazon SageMaker: Build high-quality datasets for model development workflows
- Cut generative AI costs: Code for rules, models for judgment
- Exclusive: Cogent Security's DefenseOS adds policies and automations for AI agents
- Semperis updates Lightning identity security platform to defend against AI agent attacks
- The changing fortunes of Sharon AI and the neocloud trade
- Anthropic promises to pursue humanity’s benefit. Its IPO filing gives it wiggle room
- Millionaire factory: Anthropic IPO set to fuel AI gold rush in London’s high-end property market
- Fund managers report AI efficiency gains despite training shortfall: Study
- Voyage Labs raises $12 million for a specialized AI travel assistant
- AI industry is 'bigger than one person': Cisco CEO Chuck Robbins
- SentinelOne named visionary leader in AI SOC radar
Comments
Please log in to post a comment.