Saab tests its own AI agents on Gripen E fighter for autonomous beyond visual range air combat.

Saab announced on November 3, 2026, that its in-house artificial intelligence agents autonomously controlled a Gripen E fighter jet during beyond-visual-range (BVR) air combat flight trials. Developed by Saab’s Advanced Programs division and trained across millions of simulated flight hours through its AI Fighter Pilot Academy, the proprietary software executed tactical maneuvering and target tracking. The flight campaign, built on 2025's Project Beyond to support future crewed and uncrewed combat aircraft programs, confirms the Gripen E’s software-defined avionics architecture can host autonomous flight-control applications on operational fighters without altering safety-critical flight systems.

Saab announced that its internally developed AI agents had autonomously controlled a Gripen E fighter jet during beyond-visual-range air combat flight trials, following two years of development within its Advanced Programs organization. (Picture source: Saab)

On November 3, 2026, Saab announced that its internally developed artificial intelligence (AI) agents had autonomously controlled a Gripen E fighter jet during beyond-visual-range (BVR) air combat trials, following two years of development within its Advanced Programs organization, headed by Peter Nilsson. The achievement follows three flights conducted between May 28 and June 3, 2025, under Project Beyond, when the Swedish company integrated the Centaur AI agent developed by German company Helsing into a Gripen E in less than six months. During the June 3 flight, the Centaur controlled the fighter against a human-piloted Gripen D, using live sensor information to track the opposing aircraft, execute tactical maneuvers, and provide missile firing cues to the onboard pilot.

Saab varied four conditions during the trials: initial separation, aircraft speed, relative aspect, and access to external command-and-control (C2) data, which included deliberate interruptions of C2 information. The November 2026 campaign differs because Saab developed the agents, training environments, and supporting software internally through its AI Fighter Pilot Academy, which has accumulated millions of simulated flight hours. An AI agent is a software system that receives information about its environment, evaluates possible actions, and makes decisions according to assigned objectives and operating constraints, rather than simply executing a fixed sequence of instructions.

Related topic: Saab tests AI-controlled Gripen E fighter jet against human pilot in first beyond visual range engagement

Seven categories distinguish the principal architectures of AI agents: simple reflex agents follow predefined rules; model-based agents maintain an internal representation of changing conditions; goal-based agents select actions to achieve objectives; utility-based agents compare alternatives according to performance criteria; learning agents improve through training and feedback; hierarchical agents distribute decisions between supervisory and subordinate functions; and multi-agent systems coordinate several specialized agents. In combat aviation, these functions can theoretically process radar tracks, infrared detections, aircraft speed, altitude, heading, fuel state, available weapons, and tactical data-link information.

For instance, a model-based agent could estimate an opposing fighter's position after temporary radar contact loss, while a utility-based agent could evaluate the competing risks of maintaining an engagement or executing a defensive maneuver. Reinforcement learning allows such decisions to be trained through repeated simulated engagements, allowing a software to adjust its behavior according to feedback rather than relying exclusively on programmed maneuvers. These categories differ from operational autonomy, which ranges from pilot recommendations to supervised aircraft control and independent execution of assigned tasks within predefined restrictions. The Gripen E's avionics architecture allows Saab to integrate its AI software into an operational fighter design without developing a dedicated experimental aircraft.

How Saab integrates AI agents into the Gripen E fighter jet

Powered by a F414-GE-39E engine producing 98 kN of afterburning thrust, compared with 80.5 kN for the Gripen C/D's RM12, the Gripen E measures 16.5 m in length, has an 8.6 m wingspan, a maximum takeoff weight of 16,500 kg, and 10 weapon stations supporting 7,200 kg of external stores. Its maximum speed reaches Mach 2 at altitude. The ES-05 Raven AESA radar uses a repositionable antenna providing a field of regard extending 100° to either side of the aircraft's centerline, while the Skyward-G infrared search-and-track sensor provides passive detection and Saab's Arexis system contributes electronic warfare information. These inputs, combined with tactical data links and sensor fusion, support the aircraft's tactical picture. Furthermore, the Gripen E separates mission applications from safety-critical flight-control functions, allowing AI agents to issue permitted aircraft-control commands without replacing the underlying flight-control software.

Saab integrated Helsing's Centaur in less than six months for the 2025 campaign, providing a concrete integration benchmark, while the 2026 tests now extended this approach to Saab-developed agents. Beyond-visual-range (BVR) air combat requires continuous decisions concerning aircraft geometry, sensor coverage, missile performance, and defensive maneuvering, particularly when opposing fighters operate beyond visual identification range. The Gripen E can carry the Meteor missile, whose range exceeds 100 km, and the AIM-120 AMRAAM. The Meteor's throttleable solid-fuel ducted ramjet sustains propulsion later into flight than conventional boost-coast rocket motors, improving its ability to retain energy during terminal interception. However, effective engagement distance depends on launch altitude, aircraft speed, target aspect, relative range, and subsequent maneuvers rather than nominal missile range alone.

AI-controlled maneuvering could continuously evaluate these variables while comparing offensive positioning against defensive requirements, including cranking to maintain radar coverage while increasing lateral separation, notching to complicate Doppler tracking, and turning away to increase distance from incoming missiles. Such decisions become more demanding for a pilot when several targets or threats must be managed simultaneously. Saab's 2025 trials varied speed, separation, aspect, and C2 availability to examine adaptability, while the November 2026 tests confirmed autonomous BVR control without disclosing an equivalent evaluation matrix. Training is therefore central to Saab's approach because large numbers of simulated engagements can reproduce tactical variations without consuming real fighter flight hours, fuel, engine life, or airframe fatigue margins.

How Saab trains AI fighter pilots for beyond-visual-range combat

Helsing trained its Centaur through reinforcement learning and self-play using its RL-Factory infrastructure, running thousands of concurrent engagements against simulated opponents whose behavior evolved during training. The Centaur accumulated more than 500,000 simulated flight hours, equivalent to 57.1 years of continuous aircraft operation, although parallel processing substantially reduces elapsed training time. Helsing's Vice President of Artificial Intelligence, Antoine Bordes, indicated that the version flown on June 3, 2025, accumulated the equivalent of 50 years of pilot experience within several hours of training, a comparison of simulation workload rather than human proficiency. Saab's separate AI Fighter Pilot Academy has accumulated millions of simulated hours across different air combat capabilities, evaluating decisions against defined objectives.

The principal limitation is transferring those learned behaviors into real conditions, which potentially involve radar errors, electronic warfare interference, interrupted communications, and unfamiliar opponents. For instance, Saab has not disclosed the proportion of simulated engagements involving degraded sensors and multiple threats. Sweden's investment in autonomous aviation extends to computing infrastructure, fighter research, and future uncrewed aircraft. The Swedish Defence Materiel Administration (FMV) fully financed Project Beyond under the Swedish Concept Programme for Future Fighter Systems, while Saab subsequently developed its own agents and verification tools within Advanced Programs.

Sweden's investment in AI and the development of future fighter aircraft

Saab is also a founding partner of Sferical AI, whose Sferical One installation in Linköping uses NVIDIA's DGX SuperPOD technology incorporating 1,152 GB300 GPUs, with an investment exceeding SEK 1 billion and participation from Ericsson, AstraZeneca, SEB, Wallenberg Investments, and SKF. This provides domestic computing capacity for large-scale AI training, although the hardware remains dependent on foreign semiconductor suppliers. FMV also awarded Saab SEK 2.6 billion on October 14, 2025, for future fighter research during 2025-2027, followed by SEK 2.9 billion on September 9, 2026, for activities during 2026-2028, with options extending into 2030. These contracts support broader crewed and uncrewed aircraft research rather than exclusively financing AI development.

Saab's A1, A2, and A3 research lines include the A3-001 uncrewed combat-aircraft concept unveiled at Malmen Air Base on August 22, 2026, with potential electronic warfare, suppression of enemy air defenses, and strike applications. Autonomous flying demonstrators are planned before 2030, but no A3-001 procurement quantity, unit cost, or operational service-entry date has been established. The Gripen E trials therefore demonstrate that an existing crewed fighter can incorporate AI-controlled tactical functions without becoming a fully autonomous combat aircraft. During the 2025 and 2026 campaigns, AI assumed the Gripen's maneuvering in BVR scenarios, but neither demonstrated independent missile-release authorization or complete mission execution without human supervision.

AI could progressively assume sensor-data processing, threat assessment, tactical positioning, and selected flight-control responsibilities, leaving pilots to supervise engagements and coordinate other aircraft. Uncrewed fighters also avoid human physiological constraints such as fatigue, hypoxia, and G-induced loss of consciousness, although aerodynamic, structural, propulsion, and fuel limitations remain. Comparable experimental programs include DARPA's Air Combat Evolution initiative, which used the modified F-16D X-62A VISTA for autonomous air combat maneuvering against human-controlled aircraft. The U.S. Air Force's Collaborative Combat Aircraft program is developing in parallel the General Atomics YFQ-42A and Anduril YFQ-44A Fury, while Shield AI is pursuing its X-BAT vertical-takeoff-and-landing combat aircraft concept.

Could AI-powered fighter jets eventually replace human pilots?

However, these developments indicate a shift toward human supervision of autonomous aircraft rather than the immediate replacement of fighter pilots by AI, as operational effectiveness also depends on mission authority, communications reliability, workload, and the number of uncrewed aircraft one operator can manage. The operational value of Saab's agents will depend on whether autonomous maneuvering improves engagement outcomes under conditions involving multiple opponents, electronic warfare, and disrupted communications. During BVR combat, an aircraft may simultaneously maintain radar tracking, evaluate missile opportunities, respond to incoming weapons, and coordinate with other fighters, creating competing demands on pilot attention.

AI could shorten the decision cycle by continuously evaluating relative range, closure rate, missile engagement envelopes, and defensive options, but faster processing does not automatically improve interception probability or survivability. The deliberate interruption of C2 data during the June 2025 Centaur trials addressed a specific vulnerability: maintaining tactical maneuvering when external targeting updates or command links become unavailable. Extending these functions to uncrewed aircraft could distribute sensors, weapons, and electronic warfare equipment across larger formations without requiring an onboard pilot in every aircraft.

However, effectiveness would depend on communications resilience, payload, endurance, attrition rates, procurement costs, and the number of aircraft a single Gripen E pilot could supervise. A larger number of AI-powered combat aircraft would not necessarily produce a proportional increase in combat effectiveness if their sensors were degraded, communications were interrupted, or human operators could not supervise the resulting workload.

Explore More Defense News

• Land Defense News
• Naval Defense News
• Defense Aerospace News

Written by Jérôme Brahy

Jérôme Brahy is a defense analyst and documentalist at Army Recognition. He specializes in naval modernization, aviation, drones, armored vehicles, and artillery, with a focus on strategic developments in the United States, China, Ukraine, Russia, South Korea, Türkiye, and Belgium. His analyses go beyond the facts, providing context, identifying key actors, and explaining why defense news matters on a global scale.