{"issue":{"id":"2026-10-12","coversLabel":"October 5–11, 2026","number":4,"readMinutes":6,"weekOf":"2026-10-12","publishedLabel":"Week of October 12, 2026","title":"Cheaper small models and two open-weight promises","dek":"Haiku 5.5 changes processing costs, Mistral and Reflection announce models before public weights arrive, and official UK and FINRA briefings focus on agent privacy and AI-enabled fraud.","companyLabel":"Public edition","models":[{"slug":"anthropic-haiku-5-5-tiered-pricing","kind":"model","source":{"title":"Introducing Claude Haiku 5.5","url":"https://www.anthropic.com/claude-haiku-5-5","date":"2026-10-07"},"headline":"Haiku 5.5 lowers short-request prices, while maximum-effort scores hide a trade-off","whatChanged":"Anthropic released Claude Haiku 5.5 on October 7 through its platform, AWS, Google Cloud, and Microsoft Azure. Prompts up to 100,000 tokens cost $0.10 per million input tokens and $0.50 for output, versus Haiku 4.5's $1 and $5. Above that prompt threshold, Haiku 5.5 costs $0.50 input and $2.50 output. Anthropic also halved Sonnet 5.5 cache-read pricing to $0.10 per million tokens.\n\nThe system card reports independently run Artificial Analysis GDPval-AA v2.1 results of 1620 Elo for Haiku 5.5 at max effort, versus 735 for Haiku 4.5 and 1840 for Sonnet 5.5. At Haiku's default medium effort, the rating falls to 1277 while output-token use is about a tenth of max. These are professional-work ratings, not accuracy percentages.\n\nThe launch footnote says the new tokenizer uses slightly more tokens per task. The advertised average saving therefore differs from the posted token-price reduction, and neither guarantees savings on a particular workflow.","whyItMatters":"A queue of short summaries or classification requests is a plausible place to try Haiku. Replacing a larger agent on a difficult workflow is a different decision. The gap between medium and max effort shows why a launch leaderboard cannot predict the setting your application uses.\n\nSet effort explicitly and measure cost per accepted result, including retries and corrections. Repeated agent prompts may also benefit from Sonnet's separate cache-read cut without any model substitution. Check that your requests actually generate cache hits. The migration guide is useful for API compatibility, but its living checklist supplies current integration advice rather than proof that every detail was already published at launch.","issuer":"Anthropic","glance":"Haiku 5.5 launched October 7 with tiered token prices; Sonnet 5.5 cache reads fell to $0.10 per million tokens.","keyFact":"$0.10 in, $0.50 out for prompts up to 100k","action":"Benchmark one short-request queue at medium effort and compare accepted outputs, corrections, and total cost.","attachments":[{"title":"Claude Haiku 5.5 System Card","url":"https://www.anthropic.com/claude-haiku-5-5-system-card","description":"Evaluation methods and safety results. Section 8.10.2 separates independent GDPval-AA max-effort and medium-effort ratings and token use.","source":{"title":"Claude Haiku 5.5 System Card","url":"https://www.anthropic.com/claude-haiku-5-5-system-card","date":"2026-10-07"}},{"title":"Migrating from Haiku 4.5 to Haiku 5.5","url":"https://platform.claude.com/docs/en/models/haiku-5-5/migration-guide","description":"Current integration checklist covering adaptive thinking, token recounting, request parameters, and refusals. Living documentation without a displayed publication date; linked by the October 7 announcement.","source":{"title":"Introducing Claude Haiku 5.5","url":"https://www.anthropic.com/claude-haiku-5-5","date":"2026-10-07"}}],"evidenceNote":"Anthropic's benchmark comparisons are inline HTML and SVG, inspected in the browser. The table and linked system card verify the reported ratings, but no public standalone comparison-image URL was verified. External images are decorative intro artwork, not benchmark evidence. This single-file issue cannot publish a new screenshot asset.","companyLabel":"Public edition"},{"slug":"mistral-large-4-public-preview","kind":"model","source":{"title":"Introducing Mistral Large 4","url":"https://mistral.ai/news/mistral-large-4/","date":"2026-10-06"},"headline":"Mistral Large 4 beats Medium on AutomationBench without leading the whole chart","whatChanged":"Mistral launched the public API preview of Mistral Large 4 on October 6. Public weights are promised by the end of October, not available through this announcement. The preview runs on Mistral's European infrastructure, and the company says training is still underway.\n\nIts published Artificial Analysis AutomationBench chart reports 59.9% of objectives completed without violations, versus 6.3% for Mistral Medium 3.5, 58.3% for Kimi K3, and 62.2% for GLM-5.3. The announcement describes 657 business workflows involving applications such as Gmail and Salesforce. The chart does not specify matched reasoning effort or task budgets.\n\nThe attached October 6 technical document distinguishes proprietary SaaS terms from bespoke self-deployment agreements and says it will be updated for an open-weight release. It also cautions against critical decisions without thorough validation and human oversight. A public preview, an individually negotiated deployment, and the promised public weight release are different purchasing options.","whyItMatters":"The full figure gives a more useful buying signal than a claim of frontier performance. Large 4 improves sharply over Mistral Medium 3.5 on this test and trails GLM-5.3. Equal effort and budgets are not shown, so the comparison is a reason to test a workload rather than a complete price-performance ranking.\n\nA team seeking European hosting can investigate the API preview now. A team requiring downloadable public weights has a future commitment to evaluate. Build a non-sensitive workflow test with explicit permission limits and a known result. Include unsuccessful runs and human corrections in the cost comparison, and ask for the applicable license before designing a production dependency around self-hosting.","issuer":"Mistral","glance":"Mistral opened the Large 4 API preview October 6; public weights are promised by the end of October.","keyFact":"API preview now; public weights due in October","action":"Compare Large 4 with your current model on one bounded workflow and confirm the applicable deployment license.","visuals":[{"url":"https://mistral.ai/_astro/artificial-analysis---automationbench-v3_ZUf0Vl.webp","alt":"AutomationBench chart shows Mistral Large 4 Preview at 59.9%, Medium 3.5 at 6.3%, Kimi K3 at 58.3%, and GLM-5.3 at 62.2%.","caption":"Mistral's chart attributes results to Artificial Analysis. The unit is objectives completed without violations, in percent. Large 4 Preview scores 59.9 versus Medium 3.5's 6.3 and GLM-5.3's 62.2. The figure does not disclose matched effort levels or task budgets, and business-workflow completion does not establish legal or financial accuracy. The source page was modified October 7, within this issue's window.","credit":"Mistral; results attributed to Artificial Analysis","source":{"title":"Introducing Mistral Large 4","url":"https://mistral.ai/news/mistral-large-4/","date":"2026-10-06"}}],"attachments":[{"title":"Mistral Large 4 technical documentation for downstream providers","url":"https://legal.mistral.ai/documents/Mistral%20Large%204%20-%20Technical%20Documentation%20For%20Downstream%20Providers.pdf","description":"October 6 version 1 distinguishes proprietary SaaS, partner, and bespoke self-deployment terms, explains training-data choices, and cautions about critical decisions. It says the document will change for public open weights.","source":{"title":"Mistral Large 4 technical documentation for downstream providers","url":"https://legal.mistral.ai/documents/Mistral%20Large%204%20-%20Technical%20Documentation%20For%20Downstream%20Providers.pdf","date":"2026-10-06"}}],"companyLabel":"Public edition"},{"slug":"reflection-beam-early-access","kind":"model","source":{"title":"Introducing Beam: Reflection's 501B open-weight model","url":"https://reflection.ai/blog/introducing-beam","date":"2026-10-05"},"headline":"Reflection previews Beam with competitive older-baseline scores and pending weights","whatChanged":"Reflection introduced Beam on October 5 as a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active per token. It invited early-access registrations while final red-teaming continued. Public weights, the technical report, model card, and developer artifacts were promised later in October under Apache 2.0. This was an announcement, not an immediate public download.\n\nReflection's coding table, explicitly updated October 8, puts Beam at 44.4 on DeepSWE v1.1, close to GLM 5.2's 44.0 and below GLM 5.3's 61.0. On Terminal Bench v2.1, Beam scores 80.1 versus GLM 5.2's 81.0 and GLM 5.3's 88.2. These are Reflection's published comparisons, not our independent tests.\n\nIts efficiency figures estimate generation compute from active parameters and generated tokens. They exclude prompt prefill, context-dependent attention, and serving overhead. The claim of less inference compute is not a measured hosting bill or API price.","whyItMatters":"Beam is close to GLM 5.2 on the cited tests, while the same table shows newer competitors ahead. That matters if the purchase depends on current coding performance rather than availability of public weights. The provider-selected results do not show a universal winner or matched production settings.\n\nCompanies interested in controlling their own deployment can prepare a bounded coding test while waiting for the actual artifacts. Separate total parameters from active parameters when asking a host about memory and throughput. Require measured expense per completed, checked task. The missing technical report and final red-teaming are reasons to keep the announcement out of a production replacement decision this week.","issuer":"Reflection","glance":"Reflection announced Beam October 5 with early-access registration; public weights and reports are promised later in October.","keyFact":"501B total, 23B active; weights still pending","action":"Prepare a reproducible coding task and ask for measured hosting cost when Beam artifacts become available.","visuals":[{"url":"https://reflection.ai/_next/image?url=https%3A%2F%2Fcdn.sanity.io%2Fimages%2Fsp40emik%2Fproduction%2F954e6d4da152cd6bcf8ff1d4705c06c25efc6975-2070x1206.png%3Fw%3D3840%26fit%3Dmax&w=3840&q=75","alt":"Reflection coding table compares Beam with seven open-model peers; Beam scores 44.4 on DeepSWE v1.1 and 80.1 on Terminal Bench v2.1.","caption":"Reflection's coding comparison, updated October 8. DeepSWE v1.1 scores are 44.4 for Beam, 44.0 for GLM 5.2, and 61.0 for GLM 5.3; Terminal Bench v2.1 scores are 80.1, 81.0, and 88.2. The table labels total and active parameters and prospective licenses, but does not establish equal effort, tools, or serving cost. Reflection reports these results; public Beam weights were still promised for later October.","credit":"Reflection","source":{"title":"Introducing Beam: Reflection's 501B open-weight model","url":"https://reflection.ai/blog/introducing-beam","date":"2026-10-05"}}],"evidenceNote":"The benchmark image was verified. No released Beam technical report, model card, or public weights were linked in the retrieved announcement; those artifacts remain promised for later October, so none is attached as an available document.","companyLabel":"Public edition"}],"rules":[{"slug":"ico-agentic-ai-evidence-call","kind":"rule","source":{"title":"ICO secures changes from leading AI developers as scrutiny extends to AI agents","url":"https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2026/10/ico-secures-changes-from-leading-AI-developers-as-scrutiny-extends-to-AI-agents/","date":"2026-10-08"},"headline":"The UK seeks agent evidence before writing further privacy guidance","whatChanged":"The UK's ICO opened an agentic-AI call for evidence on October 8, closing November 20. It seeks input on security, transparency, accountability, automated decisions, fairness, purpose limitations, and lawful processing. Responses will inform future guidance and a forthcoming statutory code of practice. The consultation is not a new binding rule or a completed code.\n\nThe same announcement says ten foundation-model developers have made, or committed to make, data-protection changes after supervision. These include clearer disclosures, stronger rights mechanisms, and tougher safeguard assessments. Commitments are not proof that every change has been implemented.\n\nThe ICO also confirmed ongoing enquiries with OpenAI, Anthropic, Meta, and the UK's AI Security Institute concerning agent testing and deployment. It described reported bypasses and unauthorised communication channels as potential concerns. The release does not establish findings of wrongdoing or an order to stop using those providers.","whyItMatters":"A company deploying agents can contribute a specific problem instead of a general view about AI. Document whether it can tell a user where an agent sent personal data and stop the next tool call. Identify who is responsible when a connected service changes the purpose of processing. Those checks match the call's practical focus.\n\nUse one non-sensitive workflow to check recipients, logs, and permission limits now. The ICO says policy alone cannot address these challenges, which is a reason to test technical controls rather than add another statement to a handbook. Responses may be published, so strip personal and confidential information. The November deadline is for consultation input, not a newly announced compliance deadline for every AI user.","issuer":"UK ICO","glance":"The ICO opened an agentic-AI evidence call October 8, closing November 20, to inform future guidance and a statutory code.","status":"proposed","action":"Test whether one agent workflow can list its data recipients and stop an unapproved onward transmission.","attachments":[{"title":"ICO agentic-AI evidence themes","url":"https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/the-ico-s-call-for-evidence-on-agentic-ai/","description":"Explains the evidence sought on security, transparency, accountability, automated decisions, fairness, and lawful processing. It informs future guidance rather than announcing a final code.","source":{"title":"Agentic AI call for evidence","url":"https://ico.org.uk/about-the-ico/ico-and-stakeholder-consultations/2026/10/agentic-ai-call-for-evidence","date":"2026-10-08"}},{"title":"Consultation dates and response privacy statement","url":"https://ico.org.uk/about-the-ico/ico-and-stakeholder-consultations/2026/10/agentic-ai-call-for-evidence","description":"Confirms the November 20 closing date and warns that organisational responses may be published. Useful before submitting a workflow example.","source":{"title":"Agentic AI call for evidence","url":"https://ico.org.uk/about-the-ico/ico-and-stakeholder-consultations/2026/10/agentic-ai-call-for-evidence","date":"2026-10-08"}}],"companyLabel":"Public edition"},{"slug":"finra-ai-impersonation-verification","kind":"rule","source":{"title":"How AI Is Supercharging Investment Fraud (And What You Can Do About It)","url":"https://www.finra.org/media-center/finra-unscripted/how-ai-is-supercharging-investment-fraud-and-what-you-can-do-about-it","date":"2026-10-06"},"headline":"FINRA warns that realistic impersonation leaves old trust shortcuts unreliable","whatChanged":"FINRA's October 6 educational briefing describes how AI makes impersonation, fake websites, and social engineering cheaper and more convincing. Christine Kieffer and Mayur Patel explain that a convincing video call is no longer sufficient evidence of identity and that an AI search answer can repeat a fabricated online persona. Urgency and demands for secrecy remain useful warning signs.\n\nThe episode also describes the Financial Intelligence Fusion Center as a two-way portal for member-firm cyber and fraud intelligence, plus a takedown toolkit for imposter websites. These are existing resources discussed this week, not newly launched services.\n\nFINRA expressly says the podcast creates no rule, amendment, or interpretation, and following its recommendations does not establish compliance with all applicable duties. Its transcript may contain errors. This article paraphrases the published discussion rather than presenting it as formal regulatory requirements or quoting unverified audio.","whyItMatters":"The warning applies to a supplier bank-detail change or an executive's urgent transfer request as well as an investment pitch. Do not ask the suspicious sender for the number to call. Use a contact already on file and separate the person requesting the payment from the person approving the exception.\n\nAn AI search answer is also not an identity check. A fraudster can populate the web with material that a search assistant repeats. Staff training should include a convincing message that asks for urgency or secrecy, then test whether employees pause and use an independent channel. This is an operational lesson from an educational briefing, not a newly imposed cross-industry FINRA rule or a guarantee that any single control stops every scam.","issuer":"FINRA","glance":"FINRA published October 6 educational advice on AI-enabled impersonation; the podcast creates no new rule.","status":"guidance","action":"Rehearse a plausible supplier bank-detail change and check that approval uses a contact already on file.","attachments":[{"title":"FINRA episode and published transcript","url":"https://www.finra.org/media-center/finra-unscripted/how-ai-is-supercharging-investment-fraud-and-what-you-can-do-about-it","description":"Public episode page includes practical verification discussion, member-resource explanations, and the express no-new-rule disclaimer. Transcript may contain errors; consult the audio before quoting it.","source":{"title":"How AI Is Supercharging Investment Fraud (And What You Can Do About It)","url":"https://www.finra.org/media-center/finra-unscripted/how-ai-is-supercharging-investment-fraud-and-what-you-can-do-about-it","date":"2026-10-06"}}],"companyLabel":"Public edition"}],"older":{"id":"2026-10-05","publishedLabel":"Week of October 5, 2026","title":"Cheaper models, gated releases, and new California AI laws","isLatest":false},"newer":{"id":"2026-10-19","publishedLabel":"Week of October 19, 2026","title":"Three new models, a price clock, two AI vendor probes","isLatest":true}},"session":null}