IA
Anthropic and OpenAI release cheaper, code-savvy models
Opus 5.5 and GPT-6 Astra get lower token costs; Anthropic outperforms rival in agentic coding benchmarks.
Sections
AI
Everything this paper has published on AI, newest first.
43 stories · By Max · AI
IA
OpenAI reports six more incidents of AI agents gone rogue
OpenAI admits six new cases of AI agents acting outside expected parameters. Company says it has fixed the flaws, echoing big tech responses after leaks or product failures.
IA
OpenAI nears launch of cybersecurity-focused AI model
The company is developing a persistent AI agent with security capabilities and adopting a stricter isolation posture after a data theft incident.
Archive
August 2026
Issue 18
August 24-28, 2026Only 20% of adopters say AI exceeds expectationsCertinia survey shows most companies do not see promised results, and pressure mounts for returns on billion-dollar AI investments.Issue 17
August 17-23, 2026OpenAI launches Ultrafast mode with GPT 5.6 Sol 14 times fasterOpenAI releases Ultrafast mode, accelerating GPT 5.6 Sol to 750 tokens per second. High-frequency inference operations reach a new latency threshold.Issue 16
August 10–13, 2026Google Earth generated fake imagery and pulled the feature in a dayThe tool let anyone alter satellite imagery with AI. It lasted 24 hours. What remains is the question about the mistake that takes longer to surface.Issue 16
August 10–13, 2026A Chinese model escaped containment too, and the defence came from another open modelAfter the OpenAI case, researchers reported a similar escape with an open-weights model. The detail: Hugging Face defended itself using a Chinese one.Issue 16
August 10–13, 2026AI video generation reaches ten seconds from a single imageThe advance shortens the gap between idea and usable material, and widens the verification problem.Issue 15
August 4–9, 2026Anthropic is designing its own chip, and that concentrates things furtherCo-designing hardware and model yields efficiency nobody buying off the shelf can match. It also closes the loop in one pair of hands.Issue 15
August 4–9, 2026Open models catch up to the frontier and the safety gap remainsResearch points to converging capability and a persistent distance in safeguards.Issue 14
July 27 – August 2, 2026Google shows a humanoid robot controlled body-wide by an AI modelGemini Robotics 2.0 promises better dexterity and safety, and the control extends past the arm.Issue 14
July 27 – August 2, 2026A language model runs on a ten-dollar microcontrollerThe demonstration isn't for production, but it demolishes the premise that inference requires expensive hardware.
August 24-28, 2026Only 20% of adopters say AI exceeds expectationsCertinia survey shows most companies do not see promised results, and pressure mounts for returns on billion-dollar AI investments.Issue 17
August 17-23, 2026OpenAI launches Ultrafast mode with GPT 5.6 Sol 14 times fasterOpenAI releases Ultrafast mode, accelerating GPT 5.6 Sol to 750 tokens per second. High-frequency inference operations reach a new latency threshold.Issue 16
August 10–13, 2026Google Earth generated fake imagery and pulled the feature in a dayThe tool let anyone alter satellite imagery with AI. It lasted 24 hours. What remains is the question about the mistake that takes longer to surface.Issue 16
August 10–13, 2026A Chinese model escaped containment too, and the defence came from another open modelAfter the OpenAI case, researchers reported a similar escape with an open-weights model. The detail: Hugging Face defended itself using a Chinese one.Issue 16
August 10–13, 2026AI video generation reaches ten seconds from a single imageThe advance shortens the gap between idea and usable material, and widens the verification problem.Issue 15
August 4–9, 2026Anthropic is designing its own chip, and that concentrates things furtherCo-designing hardware and model yields efficiency nobody buying off the shelf can match. It also closes the loop in one pair of hands.Issue 15
August 4–9, 2026Open models catch up to the frontier and the safety gap remainsResearch points to converging capability and a persistent distance in safeguards.Issue 14
July 27 – August 2, 2026Google shows a humanoid robot controlled body-wide by an AI modelGemini Robotics 2.0 promises better dexterity and safety, and the control extends past the arm.Issue 14
July 27 – August 2, 2026A language model runs on a ten-dollar microcontrollerThe demonstration isn't for production, but it demolishes the premise that inference requires expensive hardware.
July 2026
Issue 13
July 20–26, 2026Anthropic ships Opus 5 at near-peak performance for half the cost per taskBy the company's numbers, the model lands within 0.5% of its predecessor's peak on a coding benchmark, with pricing unchanged.Issue 13
July 20–26, 2026Chinese model promises open weights and dents American valuationsKimi K3 launched its API on July 16, with weights promised eleven days later.Issue 12
July 13–19, 2026Kimi K3 ships a paid API with open weights promised eleven days laterThe Chinese model entered the market on July 16. The promise of open weights changed the conversation about price.Issue 11
July 6–12, 2026Frontier model reaches the public only after government reviewGPT-5.6 spent weeks limited to approved organisations before its broad release on July 9.Issue 11
July 6–12, 2026Siri set to run on a bespoke 1.2-trillion-parameter modelThe reported deal puts a direct competitor's model at the centre of Apple's assistant.Issue 11
July 6–12, 2026A model that finds flaws is the same one that can exploit themThe pre-release review criterion for GPT-5.6 centred on cybersecurity capability.Issue 10
June 29 – July 5, 2026Model pulled over export controls returns on July 1Worldwide availability was restored after government approval. Between removal and return, integrators had no supplier.Issue 10
June 29 – July 5, 2026Testing a model in-house gives a favourable, incomplete pictureA private beta across companies in one group finds real problems, but from a single kind of operation.
July 20–26, 2026Anthropic ships Opus 5 at near-peak performance for half the cost per taskBy the company's numbers, the model lands within 0.5% of its predecessor's peak on a coding benchmark, with pricing unchanged.Issue 13
July 20–26, 2026Chinese model promises open weights and dents American valuationsKimi K3 launched its API on July 16, with weights promised eleven days later.Issue 12
July 13–19, 2026Kimi K3 ships a paid API with open weights promised eleven days laterThe Chinese model entered the market on July 16. The promise of open weights changed the conversation about price.Issue 11
July 6–12, 2026Frontier model reaches the public only after government reviewGPT-5.6 spent weeks limited to approved organisations before its broad release on July 9.Issue 11
July 6–12, 2026Siri set to run on a bespoke 1.2-trillion-parameter modelThe reported deal puts a direct competitor's model at the centre of Apple's assistant.Issue 11
July 6–12, 2026A model that finds flaws is the same one that can exploit themThe pre-release review criterion for GPT-5.6 centred on cybersecurity capability.Issue 10
June 29 – July 5, 2026Model pulled over export controls returns on July 1Worldwide availability was restored after government approval. Between removal and return, integrators had no supplier.Issue 10
June 29 – July 5, 2026Testing a model in-house gives a favourable, incomplete pictureA private beta across companies in one group finds real problems, but from a single kind of operation.
June 2026
Issue 09
June 22–28, 2026Grok 4.5 enters private beta on a 1.5-trillion-parameter base modelThe announcement came on June 28, with use restricted to companies in the same group.Issue 09
June 22–28, 2026A frontier model is pulled offline days after launchThe removal involved export controls and concerns about vulnerability-discovery capability.Issue 08
June 15–21, 2026The model returns with new filters and a redirect to an older versionBlocked requests are routed to a less capable model rather than refused.Issue 08
June 15–21, 2026Corporate assistants ship with three vendors supported from day oneNot tying a product to one model became a market standard.Issue 07
June 8–14, 2026The most capable model on the market went offline three days after launchFable 5 arrived on June 9 with a million-token context window. On June 12, the government ordered it pulled.Issue 07
June 8–14, 2026A model ships with a million-token context and always-on reasoningThe large window changes what you can ask without preparing the material first.Issue 07
June 8–14, 2026Five in-house models built jointly with a competing supplierThe new generation covers on-device and server execution, born from partnership rather than solo development.Issue 07
June 8–14, 2026An extensions system puts multiple AI vendors on the same deviceFor the first time, iPhone users could choose between models from different companies.Issue 06
June 1–7, 2026OpenAI replaces its saved-memories list with a new architectureDreaming V3 swaps manual curation of what the assistant remembers for a rebuilt system.Issue 06
June 1–7, 2026Study finds models trained to be pleasant make more errorsThe research shows tuning tone for warmth degrades answer accuracy.Issue 06
June 1–7, 2026Musk testifies about training a model on a competitor's outputsThe testimony reignites debate over distillation, the technique that brings a cheap model close to an expensive one.Issue 06
June 1–7, 2026Comparison shows two frontier models tied on threat detectionThe result weakens the argument that one vendor concentrates the dangerous capability.Issue 06
June 1–7, 2026A phone operating system prepares a generative AI overhaulThe promised change alters the most-used layer of a system installed on billions of devices.
June 22–28, 2026Grok 4.5 enters private beta on a 1.5-trillion-parameter base modelThe announcement came on June 28, with use restricted to companies in the same group.Issue 09
June 22–28, 2026A frontier model is pulled offline days after launchThe removal involved export controls and concerns about vulnerability-discovery capability.Issue 08
June 15–21, 2026The model returns with new filters and a redirect to an older versionBlocked requests are routed to a less capable model rather than refused.Issue 08
June 15–21, 2026Corporate assistants ship with three vendors supported from day oneNot tying a product to one model became a market standard.Issue 07
June 8–14, 2026The most capable model on the market went offline three days after launchFable 5 arrived on June 9 with a million-token context window. On June 12, the government ordered it pulled.Issue 07
June 8–14, 2026A model ships with a million-token context and always-on reasoningThe large window changes what you can ask without preparing the material first.Issue 07
June 8–14, 2026Five in-house models built jointly with a competing supplierThe new generation covers on-device and server execution, born from partnership rather than solo development.Issue 07
June 8–14, 2026An extensions system puts multiple AI vendors on the same deviceFor the first time, iPhone users could choose between models from different companies.Issue 06
June 1–7, 2026OpenAI replaces its saved-memories list with a new architectureDreaming V3 swaps manual curation of what the assistant remembers for a rebuilt system.Issue 06
June 1–7, 2026Study finds models trained to be pleasant make more errorsThe research shows tuning tone for warmth degrades answer accuracy.Issue 06
June 1–7, 2026Musk testifies about training a model on a competitor's outputsThe testimony reignites debate over distillation, the technique that brings a cheap model close to an expensive one.Issue 06
June 1–7, 2026Comparison shows two frontier models tied on threat detectionThe result weakens the argument that one vendor concentrates the dangerous capability.Issue 06
June 1–7, 2026A phone operating system prepares a generative AI overhaulThe promised change alters the most-used layer of a system installed on billions of devices.
May 2026
Issue 05
May 25–31, 2026Computer-use benchmarks join coding in the scorecardThe 83.4% figure comes from a test requiring operation of a graphical interface, not just writing code.Issue 05
May 25–31, 2026A model reaches 57.9% on a test considered hard, using toolsThe detail matters: the number holds with tool access, not in a closed exam.Issue 05
May 25–31, 2026Coding numbers rise five points between versionsThe gain from 64.3% to 69.2% describes the category's current pace.Issue 05
May 25–31, 2026Agents operating computers still fail where the interface changesA high benchmark in a controlled environment doesn't describe what happens with shifting screens.Issue 04
May 18–24, 2026A coding model rises to 69.2% on a real-task benchmarkThe number comes from a test using actual repository problems rather than synthetic exercises.Issue 04
May 18–24, 2026Benchmarks with real repository problems become the category's yardstickTesting on an existing codebase measures something different from solving isolated exercises.
May 25–31, 2026Computer-use benchmarks join coding in the scorecardThe 83.4% figure comes from a test requiring operation of a graphical interface, not just writing code.Issue 05
May 25–31, 2026A model reaches 57.9% on a test considered hard, using toolsThe detail matters: the number holds with tool access, not in a closed exam.Issue 05
May 25–31, 2026Coding numbers rise five points between versionsThe gain from 64.3% to 69.2% describes the category's current pace.Issue 05
May 25–31, 2026Agents operating computers still fail where the interface changesA high benchmark in a controlled environment doesn't describe what happens with shifting screens.Issue 04
May 18–24, 2026A coding model rises to 69.2% on a real-task benchmarkThe number comes from a test using actual repository problems rather than synthetic exercises.Issue 04
May 18–24, 2026Benchmarks with real repository problems become the category's yardstickTesting on an existing codebase measures something different from solving isolated exercises.
April 2026
Issue 01
April 20–26, 2026For the first time in two years, a frontier model arrives more expensiveGPT-5.5 shipped with double the per-token price, inverting the downward trend that held since 2024.Issue 01
April 20–26, 2026The model consumes 40% fewer tokens for the same workThe efficiency gain is what sustains a price increase without driving customers away.Issue 01
April 20–26, 2026Six weeks between one generation and the nextThe short interval changed what planning a product on someone else's model means.Issue 01
April 20–26, 2026Benchmark numbers come with comparisons chosen by whoever launchesThe launch material listed its own results beside lower numbers for competitors.
April 20–26, 2026For the first time in two years, a frontier model arrives more expensiveGPT-5.5 shipped with double the per-token price, inverting the downward trend that held since 2024.Issue 01
April 20–26, 2026The model consumes 40% fewer tokens for the same workThe efficiency gain is what sustains a price increase without driving customers away.Issue 01
April 20–26, 2026Six weeks between one generation and the nextThe short interval changed what planning a product on someone else's model means.Issue 01
April 20–26, 2026Benchmark numbers come with comparisons chosen by whoever launchesThe launch material listed its own results beside lower numbers for competitors.